A data-driven neural network prediction method for surface static stress of a runner blade

Abstract

The application discloses a data-driven runner blade surface static stress neural network prediction method, which comprises the following steps: establishing a full-flow passage model of a Francis turbine unit, performing grid division on the model, importing the grid corresponding to a working condition point into a numerical simulation software, and obtaining runner blade surface pressure data; establishing a water turbine unit shaft system model, performing non-structural grid division on the model, inputting the runner blade surface pressure data into an interface between a fluid domain and a solid domain, obtaining runner blade surface static stress data and corresponding three-dimensional coordinates, and dividing the static stress data into a training set and a test set; training a prediction model by using the three-dimensional coordinates, water head and opening degree corresponding to the training set, inputting the three-dimensional coordinates, water head and opening degree of the working condition point into the trained prediction model, and obtaining runner blade surface static stress under the working condition point. The application can correctly reflect the static stress distribution trend of the blade under all working conditions and has high precision.

Description

Technical Field

[0001] This invention relates to the field of turbine runner blade condition monitoring technology, specifically to a data-driven neural network prediction method for turbine runner blade surface static stress. Background Technology

[0002] Hydropower turbine runners are typically in a high-speed rotational state. The pressure load generated by the water flow causes the stress on the runner to change continuously. Once this alternating stress exceeds the material's bearing capacity, fatigue damage will occur. Fatigue damage usually manifests as the formation and propagation of cracks. Under prolonged stress, micro-cracks gradually form on the runner surface. These cracks, through continuous propagation and merging, eventually lead to fatigue failure of the runner. Fatigue damage to the turbine runner affects its performance and safety, potentially leading to runner fracture or even serious accidents. Therefore, real-time monitoring and assessment of the stress on the turbine runner, and taking corresponding measures to prevent fatigue damage, are crucial for ensuring the safe operation of the turbine.

[0003] Since the efficiency and accuracy of runner stress prediction directly determine the efficiency and accuracy of runner stress optimization, it is particularly important to calculate the stress on the surface of a turbine runner quickly and accurately. Traditional fluid-structure interaction calculations are too time-consuming and are gradually failing to meet practical engineering needs. For complex hydraulic machinery runners, obtaining surface stress requires a large amount of computational resources. A complete transition process typically takes tens of seconds, making it very difficult to accurately reflect its stress characteristics through numerical calculations. Exploring prediction methods that are faster and consume fewer computational resources is imperative.

[0004] A review of domestic and international research progress reveals that researchers often choose to obtain key parameters related to the lifespan of mechanical equipment through experimental measurements or finite element analysis. This data is then used to extract and organize datasets for building machine learning prediction models. By training and predicting these samples, predictive results on key influencing factors of mechanical equipment lifespan are obtained, thereby achieving mechanical equipment lifespan management. In recent years, many scholars have taken minimizing mechanical equipment lifespan loss as an optimization objective, proposing a series of models based on machine learning and deep learning technologies to predict factors affecting mechanical equipment lifespan, providing new research ideas for mechanical equipment lifespan management. Considering that neural network prediction is time-efficient and has good generalization performance, it is possible to apply neural network methods to the static stress prediction of complex hydraulic machinery blades. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a data-driven neural network prediction method for the surface static stress of turbine runner blades, and to establish a prediction model for the surface static stress of turbine runner blades based on a fully connected neural network, which can provide guidance and modification suggestions for the overall operation status of the hydropower unit.

[0006] To solve the above technical problems, the present invention adopts the following technical solution:

[0007] A data-driven neural network-based method for predicting static stress on the surface of turbine blades includes the following steps:

[0008] S1. Based on the drawings of the mixed-flow turbine unit, establish a full flow channel model of the mixed-flow turbine unit, divide the model into grids, select a set number of operating points according to the head and opening, and import the grids corresponding to the operating points into numerical simulation software for flow field calculation to obtain the surface pressure data of the runner blades.

[0009] S2. Based on the hydropower unit shaft system drawings, establish a hydropower unit shaft system model, perform unstructured mesh generation on the model, input the runner blade surface pressure data from step S1 into the interface between the fluid domain and the solid domain, perform static analysis using the finite element method, obtain the static stress data of the runner blade surface and its corresponding three-dimensional coordinates, and divide the static stress data into training set and test set.

[0010] S3. Select a prediction model using the three-dimensional coordinates, head, and opening of any operating point. Train the prediction model using the corresponding three-dimensional coordinates, head, and opening of the training set. Input the three-dimensional coordinates, head, and opening of the operating point into the trained prediction model to obtain the static stress on the surface of each impeller blade at that operating point.

[0011] Furthermore, in step S1, obtaining the surface pressure data of the impeller blades includes the following sub-steps:

[0012] S101. Based on the drawings of the mixed-flow turbine unit, a full flow channel model of the mixed-flow turbine unit is established using SolidWorks software. The model includes the volute, fixed guide vanes, movable guide vanes, runner, pressure reducing pipe, and tailrace pipe.

[0013] S102. The mixed-flow turbine model is meshed using n different mesh number schemes. Under the same boundary conditions, steady-state fluid numerical simulation is performed on the mixed-flow turbine model based on rated head and rated power to obtain the corresponding runner efficiency. The scheme with the highest runner efficiency and the fewest mesh numbers is selected as the final meshing scheme. The runner is divided into unstructured meshes, and the rest is divided into structured meshes.

[0014] S103. Based on the pressure head and the opening of the movable guide vane corresponding to different operating conditions, the intersection of the equal head line and the equal efficiency line on the comprehensive characteristic curve of the entire flow channel of the mixed-flow turbine unit is set as the head opening, and a set number of operating points are selected according to the head opening.

[0015] S104. Use ICEM software to re-mesh the active guide vanes with different openings under different working conditions to obtain a hexahedral structure mesh. Other components are not meshed. Assemble all the meshes so that each working point corresponds to an independent mesh.

[0016] S105. Use the Fluent solver to perform computational fluid dynamics analysis on the mesh from step S104 to obtain the surface pressure data of the runner blades. The specific details are as follows:

[0017] Fluid motion satisfies the laws of mass conservation and energy conservation, and the corresponding equations are shown in formula (7):

[0018]

[0019] Where ρ represents density, t represents time, and u represents streamline. Let p represent the Hamiltonian operator, v represent the static pressure, and v represent the kinematic viscosity. denoted by Laplace operator, f represents the mass force.

[0020] The Navier-Stokes equations are time-averaged using the Reynolds averaging method, yielding the time-averaged Reynolds equation, specifically expressed as follows:

[0021]

[0022] Where Δt represents unit time, φ represents the time average, and φ represents the instantaneous physical quantity.

[0023] Decomposing the instantaneous physical quantity yields the instantaneous equation, the specific expression of which is:

[0024]

[0025] Where φ′ represents the pulsation value.

[0026] By simplifying equation (9) using the Cartesian tensor, we obtain the Reynolds time-averaged Navier-Stokes equations, specifically expressed as follows:

[0027]

[0028] Among them, u i Let u represent the i-th streamline velocity. j Let f represent the j-th streamline velocity. i Let u represent the i-th mass force. i ′ represents the fluctuating value of the i-th streamline velocity, u j ′ represents the fluctuating value of the j-th streamline velocity. Indicate u i ′ and u jThe time mean of the product, (x) i ,x j ) represents spatial coordinates.

[0029] The Reynolds-averaged Navier-Stokes equation is closed using the Realizable k-ε transport equation, and the specific expression is shown in formula (11):

[0030]

[0031] in, k represents turbulent kinetic energy, ε represents turbulent kinetic energy dissipation rate, and μ t G represents turbulent viscosity. k G represents the turbulent kinetic energy generated by the average velocity gradient. b Y represents the turbulent kinetic energy generated under the action of buoyancy. M C represents the portion of the total dissipation rate of compressible turbulence caused by fluctuation expansion. 1ε C2, C 3ε Both represent constants, σ k σ ε Let S represent the Prandtl numbers for turbulence at k and ε, respectively. k and S ε Both represent custom source terms; μ represents dynamic viscosity, and S represents user-defined parameters.

[0032] The runner region was processed using the multiple reference frame method, and then steady-state calculations were performed. The Reynolds-averaged Navier-Stokes equations were solved using the SIMPLEC method to obtain the velocity and pressure distributions inside the flow field, including the pressure data on the runner blade surface.

[0033] Furthermore, in step S2, obtaining the static stress data on the surface of the runner blades includes the following sub-steps:

[0034] S201. Based on the hydropower unit shaft system drawings, a hydropower unit shaft system model is established using SolidWorks software. The model includes a mixed-flow impeller, thrust bearing, generator rotor, upper guide bearing, water guide bearing, lower guide bearing, and main shaft. The surface of the blades in contact with water is set as the interface between the fluid domain and the solid domain.

[0035] S202. Using the static structure finite element software, the solid domain is divided into an unstructured mesh, and the mesh of the runner blade is refined based on the calculation accuracy.

[0036] S203. Constraints are set for the solid domain using the physical properties of the hydropower unit shaft system model. The water guide bearing, upper guide bearing, and lower guide bearing are fixed to constrain radial vibration, and the thrust bearing is fixed to constrain axial vibration. The rotational speed and gravitational acceleration are also set.

[0037] S204. Input the surface pressure data of the runner blades as a load into the interface between the fluid and solid domains to perform fluid-structure interaction calculations and obtain the static stress data of the runner blade surface. The specific content is as follows:

[0038] The governing equations ensure that the stress and displacement between the fluid and solid domains satisfy the conservation conditions. The specific formula is as follows:

[0039]

[0040] Where, τ f n f τ represents the normal stress of the fluid. s n s d represents the normal stress in a solid. f d represents fluid displacement. s Indicates solid displacement.

[0041] Finite element analysis was performed on the interface between the fluid and solid domains using a one-way steady-state fluid-structure interaction method to obtain static stress data and corresponding three-dimensional coordinates on the surface of the runner blade.

[0042] Furthermore, in step S3, predicting the static stress on the runner blade surface includes the following sub-steps:

[0043] S301. Input the three-dimensional coordinates, head, and opening of the surface mesh node of any blade of the runner at any operating point into each neural network model to obtain the predicted static stress data. Compare the error between the predicted static stress data and the test set, and select the neural network model with the error below the set error as the prediction model. The model includes an input layer, a fully connected layer, and an output layer.

[0044] S302. Input the three-dimensional coordinates, head, and opening corresponding to the training set into the prediction model. After partial differential operations and nonlinear activation function processing, obtain the weights and bias parameters.

[0045] S303. Based on the weights and bias parameters in step S302, the prediction model is trained and its parameters are adjusted using the backpropagation algorithm and gradient descent algorithm until the error between the predicted static stress data and the test set reaches below the set error threshold. Training is then stopped, and the trained prediction model is obtained.

[0046] S304. Input the three-dimensional coordinates, head, and opening of the operating point into the trained prediction model to obtain the static stress on the surface of each impeller blade at that operating point.

[0047] Furthermore, the present invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the data-driven neural network prediction method for surface static stress of turbine blades described above.

[0048] Furthermore, the present invention also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the aforementioned data-driven neural network prediction method for static stress on the surface of turbine blades.

[0049] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0050] 1. The neural network used in this invention has a rapid response in predicting the static stress of turbine blades, and can output results in just a few seconds. It can also accurately reflect the distribution trend of static stress on the blades under all operating conditions. The absolute error of the static stress on the runner blades at the test operating point is basically controlled within 1MPa, and the relative error is basically controlled below 10%, which meets the engineering requirements.

[0051] 2. The method proposed in this invention has good generalization performance and can make effective predictions for different operating points within the operating range, providing guidance and modification suggestions for the overall operation status of hydropower units. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating the overall implementation of the present invention.

[0053] Figure 2 This is a schematic diagram of the full flow path model of the mixed-flow turbine unit of the present invention.

[0054] Figure 3 This is a schematic diagram of the full flow channel mesh of the mixed-flow turbine unit model of the present invention.

[0055] Figure 4 This is a schematic diagram illustrating the mesh independence verification of the present invention.

[0056] Figure 5 This is a schematic diagram of the head output distribution at the operating point of the present invention.

[0057] Figure 6 This is a schematic diagram of the hydroelectric generator shaft system model, constraint settings, and solid domain mesh of the present invention.

[0058] Figure 7 This is a schematic diagram of the fully connected neural network model architecture of the present invention.

[0059] Figure 8 This is a schematic diagram of the head and opening distribution of the training working point and the selection of the test working point in this invention.

[0060] Figure 9 This is a schematic diagram comparing the results of the neural network prediction of static stress on the pressure and suction surfaces of a single blade with the results of finite element calculation.

[0061] Figure 10 This is a schematic diagram comparing the neural network prediction results and finite element calculation results of all blade static stresses in this invention. Detailed Implementation

[0062] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0063] To achieve the above objectives, this invention proposes a data-driven neural network-based method for predicting static stress on the surface of turbine blades, such as... Figure 1 As shown, the specific steps are as follows:

[0064] S1. Based on the drawings of the mixed-flow turbine unit, a full-flow channel model of the mixed-flow turbine unit was established using SolidWorks software. This model was then meshed, and 70 operating points were selected according to the head opening. The mesh corresponding to these 70 operating points was imported into numerical simulation software for flow field calculation to obtain the surface pressure data of the runner blades. The specific content is as follows:

[0065] S101, such as Figure 2 As shown, the full flow path model of the mixed-flow turbine unit includes the volute, fixed guide vanes, movable guide vanes, runner, pressure reducing pipe, and tailrace pipe. Figure 2 The model of the entire flow channel of the mixed-flow turbine unit is displayed, with each component distinguished by a different color, showcasing the structure of the fluid computation domain.

[0066] S102, such as Figure 3 As shown, the entire flow channel model of the mixed-flow turbine unit was meshed. The runner was divided into an unstructured mesh, while the remaining parts were divided into structured meshes. Mesh generation divided the model into many small elements to ensure that numerical calculations could proceed normally and obtain accurate results. Mesh generation completed the preparation for mesh independence verification and selection of the optimal mesh.

[0067] In fluid numerical simulations, the quality of the mesh directly impacts the flow field calculation results. The number of mesh elements is equally important; too few mesh elements make it difficult to capture flow details or miss some characteristics of the flow components, while too many mesh elements require significant computational resources, slowing down the calculation. An appropriate mesh number ensures simulation accuracy and a shorter computation time, while simultaneously ensuring that the mesh does not affect the correctness and accuracy of the numerical simulation. Therefore, mesh independence verification is necessary.

[0068] like Figure 4 As shown, six mesh generation schemes were used, with total mesh counts of 3.62 million, 4.98 million, 6.41 million, 7.84 million, 10.32 million, and 13.5 million respectively. All six schemes used the same boundary conditions. Steady-state fluid numerical simulations were performed on the full-flow-channel model of the mixed-flow turbine unit based on rated head and rated power to obtain the corresponding runner efficiencies. The mesh generation scheme with the highest efficiency and the fewest meshes was selected as the final mesh generation scheme. Figure 4 As can be seen from the data, the fifth scheme with a grid size of 10.32 million has a moderate number of grids and high operating efficiency, so it is selected as the optimal grid for subsequent calculations.

[0069] S103, such as Figure 5 As shown, based on the pressure head and the opening of the movable guide vane corresponding to different operating conditions, the intersection of the equal head line and the equal efficiency line on the comprehensive characteristic curve of the entire flow channel of the mixed-flow turbine unit is set as the head opening, and 70 operating points are selected according to the head opening. Figure 5 The diagram displays the power and opening data corresponding to each operating condition, providing a basis for modifying the active guide vane grid under different operating conditions. The operating points in this diagram are typical operating points selected for hydropower stations.

[0070] S104. Since the operating conditions of the entire flow channel of the mixed-flow turbine unit need to be changed to change the opening of the movable guide vanes, and once the opening of the movable guide vanes changes, the shape of the movable guide vanes will change, and the mesh of the movable guide vanes needs to be re-divided. Therefore, ICEM software is used to re-structure the mesh of the movable guide vanes with different openings under different operating conditions to obtain a hexahedral structure mesh. Other components are not meshed. All meshes are assembled so that each different operating point corresponds to an independent mesh, for a total of 70 meshes.

[0071] S105. Using the Fluent solver, perform CFD (Computational Fluid Dynamics) analysis on the 70 mesh sets from step S104 to obtain the surface pressure data of the runner blades. The specific details are as follows:

[0072] Fluid motion must conform to the laws of conservation of mass, momentum, and energy. Applying these conservation laws to fluid motion and undergoing corresponding transformations, the Navier-Stokes (NS) equations of fluid mechanics are formed. In fluid mechanics, it is generally believed that when solving for incompressible fluids, the energy conservation equations can be disregarded due to the relatively small amount of heat exchange. Therefore, the fluid motion only needs to satisfy the conservation of mass and energy, and the corresponding equations are shown in formula (13):

[0073] The expressions for the mass conservation equation and the momentum conservation equation are as follows:

[0074]

[0075] Where ρ represents density, t represents time, and u represents streamline. Let p represent the Hamiltonian operator, v represent the static pressure, and v represent the kinematic viscosity. denoted by Laplace operator, f represents the mass force.

[0076] When solving turbulent flow problems, since the Navier-Stokes equations describing turbulent flow are difficult to find analytical solutions, the Reynolds averaging method is used to treat turbulence as the superposition of two flows, thus time-averaging the Navier-Stokes equations and obtaining the time-averaged Reynolds equations, specifically expressed as follows:

[0077]

[0078] Where Δt represents unit time, φ represents the time average, and φ represents the instantaneous physical quantity.

[0079] Decomposing the instantaneous physical quantity yields the instantaneous equation, the specific expression of which is:

[0080]

[0081] Where φ′ represents the pulsation value.

[0082] Simplifying equation (15) using the Cartesian tensor, we obtain the RANS (Reynolds Average Navier-Stokes, Reynolds Time-Average NS) equation system, with the following specific expression:

[0083]

[0084] Among them, u i Let u represent the i-th streamline velocity. j Let f represent the j-th streamline velocity. i Let u represent the i-th mass force. i ′ represents the fluctuating value of the i-th streamline velocity, u j ′ represents the fluctuating value of the j-th streamline velocity. Indicate u i ′ and u j The time mean of the product, (x) i ,x j ) represents spatial coordinates.

[0085] Since the equation system is no longer closed, it is necessary to re-establish the relationship between the time-averaged and fluctuating values. Therefore, a suitable turbulent transport equation is introduced as a new expression to close the RANS equation system, and then the solution is obtained. The Realizable k-ε transport equation is used to close the RANS equation system, and the specific expression is shown in formula (17):

[0086]

[0087] in, k represents turbulent kinetic energy, ε represents turbulent kinetic energy dissipation rate, and μ t G represents turbulent viscosity. k G represents the turbulent kinetic energy generated by the average velocity gradient. b Y represents the turbulent kinetic energy generated under the action of buoyancy. M C represents the portion of the total dissipation rate of compressible turbulent flow caused by wave expansion. 1ε C2, C 3ε Both represent constants, σ k σ ε Let S represent the Prandtl numbers for turbulence at k and ε, respectively. k and S ε All of these represent custom source terms, μ represents dynamic viscosity, and S represents user-defined parameters.

[0088] The runner region was processed using the Multiple Reference Frame (MRF) method, and steady-state calculations were performed simultaneously. The RANS equations were solved using the SIMPLEC method to obtain the velocity and pressure distributions within the flow field, including the pressure data on the blade surfaces. This data was saved in the software as a .dat file, and the software exported the blade surface pressure data as a .txt file for subsequent finite element analysis.

[0089] S2. Based on the hydropower unit shaft system drawings, a hydropower unit shaft system model is established using SolidWorks software. This model is then imported into Static Structures software for unstructured mesh generation. The surface pressure data of the turbine blades from step S1 is input into the interface between the fluid and solid domains. Static analysis is performed using the finite element method to obtain the static stress data of the turbine blade surface and its corresponding three-dimensional coordinates. This static stress data is then divided into training and testing sets. The specific content is as follows:

[0090] S201. The hydroelectric generator shaft system model includes a mixed-flow impeller, thrust bearing, generator rotor, upper guide bearing, water guide bearing, lower guide bearing, and main shaft. The surface of the blades in contact with the water is defined as the interface between the fluid domain and the solid domain.

[0091] S202. Using the static structure finite element software, the solid domain is divided into an unstructured mesh, and the mesh of the runner blade is refined based on the calculation accuracy.

[0092] S203. For the hydropower unit shaft system model, the mixed-flow runner is the most important flow component. Set the physical properties of the hydropower unit shaft system model (including material density, Young's modulus, Poisson's ratio, bulk modulus, shear modulus, tensile and compressive yield strength, and tensile ultimate strength), and set constraints on the solid domain. Fixed the radial vibration of the water guide bearing, upper guide bearing, and lower guide bearing, and fixed the axial vibration of the thrust bearing. At the same time, set the rotational speed and gravitational acceleration.

[0093] Figure 6 A schematic diagram of the shaft system model and constraint settings for the hydroelectric generator, and the solid domain mesh. Figure 6 (a) shows the shaft system model and constraint settings for the hydroelectric generator unit. Figure 6 (b) is a schematic diagram of the solid domain mesh. It can be seen that the constraint settings can make the hydropower unit shaft system model closer to the actual operation. The solid domain mesh divides the hydropower unit shaft system into sufficiently small units to facilitate subsequent fluid-structure interaction calculations.

[0094] S204. Input the surface pressure data of the runner blades as a load into the interface between the fluid and solid domains to perform fluid-structure interaction calculations and obtain the static stress data of the runner blade surface. The specific content is as follows:

[0095] In fluid-structure interaction calculations, to ensure accurate and reliable results, it is crucial to guarantee that the transfer of physical quantities between the fluid and solid domains strictly adheres to conservation laws. Specifically, the temperature, heat flux, stress, and displacement parameters at the interface between the fluid and solid domains must remain consistent. Since the calculations in this invention do not consider the effects of heat exchange, it is only necessary to ensure that stress and displacement satisfy the conservation conditions. The governing equations are used to ensure that the stress and displacement between the fluid and solid domains satisfy the conservation conditions; the specific formula is as follows:

[0096]

[0097] Where, τ f n f τ represents the normal stress of the fluid. s n s d represents the normal stress in a solid. f d represents fluid displacement. s Indicates solid displacement.

[0098] Finite element analysis was performed on the interface between the fluid and solid domains using a one-way steady-state fluid-structure interaction method to obtain static stress data and corresponding three-dimensional coordinates on the surface of the runner blade.

[0099] The static stress distribution is reflected in the visualization results of the static structure finite element software, which exports the static stress distribution on the blade surface and the static stress data of the runner blade surface at each grid point, along with their corresponding three-dimensional coordinates.

[0100] S3. Select a prediction model using the three-dimensional coordinates, head, and opening of any operating point. Train the prediction model using the corresponding three-dimensional coordinates, head, and opening from the training set. Input the three-dimensional coordinates, head, and opening of the operating point into the trained prediction model to obtain the static stress on the surface of each impeller blade at that operating point. The specific content is as follows:

[0101] S301, such as Figure 7 As shown, the three-dimensional coordinates (i.e., x1 to x5 are the three-dimensional coordinates x, y, z of each grid node on the specified blade), head, and opening of any blade at any operating point are input into each neural network model to obtain the predicted static stress data SS. The error between the predicted static stress data and the test set is compared, and the neural network model with an error of less than 10% is selected as the prediction model. In this embodiment, a fully connected neural network model is selected as the prediction model. This model includes an input layer, a fully connected layer, and an output layer, where the l-th layer is a fully connected layer (or a linear layer). A fully connected neural network contains many fully connected layers, which are briefly explained here. The e-th layer is the output layer. eΩ ζ represents the result of the nonlinear activation function operation corresponding to the Ω-th neuron in the e-th layer. eΩ Let ζ represent the result of the loss function operation corresponding to the Ω-th neuron in the e-th layer, where a represents the non-linear activation function and ζ represents the loss function.

[0102] S302. Input the three-dimensional coordinates, head, and opening corresponding to the training set into the prediction model. After partial differential operations and nonlinear activation function processing, obtain the weights and bias parameters.

[0103] S303. Based on the weights and bias parameters in step S302, the prediction model is trained and its parameters are adjusted using the backpropagation and gradient descent algorithms. Hyperparameters such as the number of training epochs, learning rate, and batch size are adjusted during training until the error between the predicted static stress data and the test set reaches below 0.0001, at which point training stops, resulting in a trained prediction model. By minimizing the loss function to adjust the network's weights and bias parameters, the model can more accurately fit the training data and exhibit good generalization and nonlinear solution capabilities.

[0104] The backpropagation algorithm is based on the chain rule, calculating the gradient of the parameters of each layer backward, starting from the output layer. First, it calculates the error of the output layer, which is the partial derivative of the loss function with respect to the output of the neurons in that layer. Then, by backpropagating these errors, it calculates the error of each neuron in each layer, i.e., the partial derivative of the loss function with respect to the output of that neuron; these errors are further used to calculate the gradient of the loss function with respect to the model parameters.

[0105] (1) Update all parameters:

[0106] For the output layer node, the error E is y e The function, and y e This represents the output of the output layer, which is also the value of the non-linear activation function net. e The function, net e It is the weight w le and bias b e The function.

[0107]

[0108] Where γ represents the learning rate, Δw le Δb represents the weight gradient from the fully connected layer to the output layer. e w represents the bias gradient of the output layer. le b represents the weights from the fully connected layer to the output layer. e This indicates the bias of the output layer.

[0109] (2) For fully connected layer nodes:

[0110]

[0111] Where, Δw ol This represents the weight gradient from the input layer to the fully connected layer. w represents the derivative of the nonlinear activation function in the fully connected layer, M represents the number of nodes in each layer of the neural network, and w ql The m represents the weight from the q-th node of the input layer to the fully connected layer. q Let y represent the value of the q-th training sample. q This represents the q-th output, Δθ l G′(net) represents the bias gradient of the fully connected layer. q ) represents the derivative of the nonlinear activation function in the output layer, and xo represents the input of the input layer.

[0112] S304. Input the three-dimensional coordinates, head, and opening of the operating point into the trained prediction model to obtain the static stress on the surface of each impeller blade at that operating point.

[0113] In this embodiment, two operating points are selected at the center and two at each end of the operating range, forming the first set of test operating points, used to analyze the prediction performance of the neural network on both sides and in the middle of the operating range; and four operating points are selected in a relatively open area, forming the second set of test operating points, used to analyze the performance of the neural network in areas where the operating points are sparsely distributed. The distribution of head opening at the training operating points and the selection of the test operating points are as follows. Figure 8 As shown. After selecting the appropriate test conditions, test calculations can be performed to verify the predictive performance of the fully connected neural network.

[0114] Visualizations are post-processed and rendered by tecplot to observe the differences between the learning and simulation results. Figure 9 This diagram illustrates the comparison between neural network predictions and finite element calculations of static stress on the pressure and suction surfaces of a single blade. Figure 9 (a) shows the neural network prediction results of static stress on the pressure surface of a single blade. Figure 9 (b) shows the finite element calculation results of static stress on the pressure surface of a single blade. Figure 9 (c) represents the neural network prediction result of static stress on the suction surface of a single blade. Figure 9 (d) represents the finite element calculation result of the static stress on the suction surface of a single blade. The comparison shows that the fully connected neural network can make predictions that meet the accuracy requirements for the static stress on both the suction and pressure surfaces of a single blade. Figure 10 This is a schematic diagram comparing the neural network prediction results and finite element calculation results for the static stress of all blades. Figure 10 (a) represents the neural network prediction results for the static stress of all blades. Figure 10 (b) shows the finite element calculation results for the static stress of all blades. The areas of static stress concentration on the runner blades are highlighted and magnified to observe the accuracy of the neural network predictions. Figure 10 It can be seen that a fully connected neural network can predict the surface static stress of all runner blades with the required accuracy.

[0115] This invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. It should be noted that when the processor executes the computer program, it corresponds to the specific steps of the method provided in this invention, possessing the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in this invention.

[0116] This invention also proposes a computer-readable storage medium storing a computer program. It should be noted that when the computer program is executed by a processor, it corresponds to the specific steps of the method provided in this invention, possessing the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in this invention.

[0117] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A data-driven based neural network prediction method for surface static stress of a runner blade, characterized in that, The method comprises the following steps: S1, based on the Francis turbine drawing, a Francis turbine full passage model is established, the model is meshed, a certain number of working condition points are selected according to the water head opening, the mesh corresponding to the working condition points is imported into the numerical simulation software for flow field calculation, and the runner blade surface pressure data is obtained; S2, based on the hydroelectric generator shafting drawing, a hydroelectric generator shafting model is established, the model is meshed, the runner blade surface pressure data in step S1 is input into the interface between the fluid domain and the solid domain, the finite element method is used for statics analysis, the runner blade surface static stress data and the corresponding three-dimensional coordinates are obtained, and the static stress data is divided into a training set and a test set; Specifically: S201, based on the hydroelectric generator shafting drawing, a hydroelectric generator shafting model is established by using SolidWorks software, the model comprises a Francis runner, a thrust bearing, a generator rotor, an upper guide bearing, a water guide bearing, a lower guide bearing and a main shaft; the surface of the blade in contact with water is set as the interface between the fluid domain and the solid domain; S202, the solid domain is divided into non-structural grids by using the static structure finite element software, and the runner blade is meshed according to the calculation accuracy; S203, the physical properties of the hydroelectric generator shafting model are used to set constraints for the solid domain, the water guide bearing, the upper guide bearing and the lower guide bearing are fixedly constrained in the radial direction, the thrust bearing is fixedly constrained in the axial direction, and the rotational speed and the gravitational acceleration are set; S204, the runner blade surface pressure data is input as a load into the interface between the fluid domain and the solid domain, fluid-structure coupling calculation is performed, and runner blade surface static stress data is obtained, and the specific content is: The control equation is used to make the stress and displacement between the fluid domain and the solid domain satisfy the conservation condition, and the specific formula is: ； wherein, represents the fluid normal stress, represents the solid normal stress, represents the fluid displacement, represents the solid displacement; The one-way steady fluid-structure coupling method is used for finite element analysis of the interface between the fluid domain and the solid domain, and the runner blade surface static stress data and the corresponding three-dimensional coordinates are obtained; S3, a prediction model is selected by using the three-dimensional coordinates of any working condition point and the water head opening, the prediction model is trained by using the three-dimensional coordinates and the water head opening corresponding to the training set, the three-dimensional coordinates and the water head opening of the working condition point are input into the trained prediction model, and the runner blade surface static stress of the working condition point is obtained.

2. The data-driven based neural network prediction method of surface static stress of a runner blade according to claim 1, wherein, In step S1, the runner blade surface pressure data comprises the following substeps: S101, based on the Francis turbine drawing, a Francis turbine full passage model is established by using SolidWorks software, the model comprises a spiral case, a fixed guide vane, a movable guide vane, a runner, a pressure reducing pipe and a tailrace pipe; S102, the Francis turbine full passage model is meshed by adopting n grid quantity schemes, under the same boundary conditions, the Francis turbine full passage model is subjected to steady fluid numerical simulation based on the rated water head and the rated power, the corresponding runner efficiency is obtained, the scheme with the highest runner efficiency and the least grid quantity is selected as the final meshing scheme, the runner is divided into non-structural grids, and the remaining parts are divided into structural grids; S103, based on the pressure head corresponding to different working conditions and the opening of the movable guide vane, set the intersection of the equal head line and the equal efficiency line on the comprehensive characteristic curve of the Francis turbine unit as the head opening, and select a certain number of working condition points according to the head and the opening; S104, using ICEM software to redivide the grid of the movable guide vane at different openings under different working conditions, get the hexahedral structure grid, other components are not divided, assemble all the grids, so that each working condition point corresponds to a set of independent grids; S105, using the fluent solver to perform computational fluid dynamics analysis on the grid in step S104, get the runner blade surface pressure data, the specific content is: The fluid motion satisfies the conservation of mass and energy, and the corresponding equation is shown in formula (1): ； where p denotes density, t denotes time, u denotes flow line, denotes the Hamiltonian operator, p denotes static pressure, v denotes kinematic viscosity, denotes the Laplace operator, f denotes mass force; The Reynolds average method is used to time-average the Navier-Stokes equation to obtain the time-average Reynolds equation, the specific expression is: ； wherein, represents the unit time, represents the time average, represents the instantaneous physical quantity; The instantaneous physical quantity is decomposed to obtain the instantaneous equation, the specific expression is: ； wherein represents a pulsation value; The formula (3) is simplified by using the Cartesian tensor to obtain the Reynolds time-average N-S equation group, the specific expression is: ； ； wherein represents the i-th streamwise velocity, represents the j-th streamwise velocity, represents the i-th mass force, represents the fluctuation of the i-th streamwise velocity, represents the fluctuation of the j-th streamwise velocity, represents and the time-averaged value of the product, (x i , x j ) represents the spatial coordinates; The Realizable k-ε transport equation is used to close the Reynolds averaged Navier-Stokes equation, the specific expression is formula (5): （5）； wherein, , k represents turbulent kinetic energy, ε represents turbulent kinetic energy dissipation rate, μ t represents turbulent viscosity, G k represents turbulent kinetic energy generated by the presence of average velocity gradient, G b represents turbulent kinetic energy generated by the action of buoyancy, Y M represents the part of the total dissipation rate of compressible turbulence caused by fluctuation expansion, , C2、 all represent constants, , respectively represent turbulent Prandtl number of k, ε, and all represent user-defined source terms; represents dynamic viscosity, represents user-defined parameters; The runner area is processed by using the multiple reference frame method, and then the steady calculation is performed, the SIMPLEC method is used to solve the Reynolds averaged Navier-Stokes equation to obtain the velocity distribution and pressure distribution inside the flow field, which includes the runner blade surface pressure data.

3. The data-driven based neural network prediction method of surface static stress of a runner blade according to claim 1, wherein, In step S3, predicting the runner blade surface static stress includes the following substeps: S301, input the three-dimensional coordinates, head and opening corresponding to the surface grid nodes of any blade of the runner under any working condition point into each neural network model to obtain the predicted static stress data, compare the error between the predicted static stress data and the test set, select the neural network model corresponding to the error below the set error as the prediction model, which includes the input layer, the full connection layer and the output layer; S302, input the three-dimensional coordinates, head and opening corresponding to the training set into the prediction model, and process through partial differential operation and nonlinear activation function to obtain the weight and bias parameters; S303, based on the weight and bias parameters in step S302, use the back propagation algorithm and the gradient descent algorithm to train and parameter adjust the prediction model, until the error between the predicted static stress data and the test set reaches below the set error threshold, stop training, and obtain the trained prediction model; S304, input the three-dimensional coordinates, head and opening of the working condition point into the trained prediction model to obtain the static stress of each runner blade surface under the working condition point.

4. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to realize the steps of the data-driven runner blade surface static stress neural network prediction method in any one of claims 1 to 3.

5. A computer-readable storage medium storing a computer program, the computer-readable storage medium being characterized by, The computer program is run by the processor to execute the data-driven runner blade surface static stress neural network prediction method in any one of claims 1 to 3.

Images