Flow field prediction method for quasi-rectangular section of pump station flow channel based on hybrid neural network

Abstract

The invention discloses a hybrid neural network-based flow field prediction method for a quasi-rectangular section of a pump station flow channel, and belongs to the technical field of pump station flow channel flow field prediction, and the method comprises the steps: constructing a scaling model pump station according to an actual pump station; based on the model pump station, a particle image velocity measurement experiment is carried out, and flow field data and working condition characteristic parameters of the quasi-rectangular cross section of the elbow-shaped water inlet flow channel under different working conditions are obtained; an MLP-PINN hybrid neural network model is constructed; training an MLP-PINN hybrid neural network model based on the flow field data and the working condition characteristic parameters; and based on the trained MLP-PINN hybrid neural network model, predicting the flow field characteristics of the quasi-rectangular section of the pump station elbow-type inlet passage under the unknown working condition. According to the method, on the basis of experimental data of part of working conditions, the flow field characteristics of other working conditions are predicted, the problems of incomplete working condition coverage and high cost are solved, more comprehensive monitoring data are provided for pump station monitoring, and the safety of the pump station is enhanced.

Description

Technical Field

[0001] This application relates to the technical field of predicting the flow field of pump station flow channels, and particularly to a method for predicting the flow field of a quasi-rectangular section of a pump station flow channel based on a hybrid neural network. Background Art

[0002] As a core facility in the cross-basin water transfer, agricultural irrigation, and urban water supply and drainage systems, the operation stability of a pump station is directly related to the project benefits and safety reliability. In axial-flow pump stations, the elbow-shaped intake channel is widely used because it can significantly optimize the water flow conditions. The elbow-shaped intake channel smoothly guides the water flow direction through its curved channel shape, avoiding turbulence, vortices, and energy losses caused by sharp turns or sudden changes in the flow direction, thereby improving the flow pattern at the pump inlet and enhancing the hydraulic efficiency. Therefore, the flow field characteristics inside the elbow-shaped channel have a decisive impact on the operation state of the pump and the overall energy efficiency of the pump station.

[0003] However, in actual projects, the operating conditions of pump stations are complex and variable, and the available real operating condition data for monitoring is extremely limited. Existing technologies rely on building a geometrically scaled model of the pump station for experiments, simulating and analyzing the flow field characteristics in the laboratory. However, model processing, experimental bench setup, and repeated testing require a large amount of capital and time, and are limited by experimental resources, making it difficult to obtain the flow field characteristics under all operating conditions. Summary of the Invention

[0004] Aiming at the above deficiencies in the prior art, a method for predicting the flow field of a quasi-rectangular section of a pump station flow channel based on a hybrid neural network provided by this application solves the problems of limited experimental resources, incomplete experimental condition coverage, and high cost in existing pump stations.

[0005] To achieve the above invention objective, the technical solution adopted by this application is as follows: This application provides a method for predicting the flow field of a quasi-rectangular section of a pump station flow channel based on a hybrid neural network, including: S1: Construct a scaled model pump station according to the actual pump station; S2: Based on the model pump station, conduct particle image velocimetry experiments, and obtain the flow field data and operating condition characteristic parameters of the quasi-rectangular section of the elbow-shaped intake channel under different operating conditions; S3: Construct an MLP-PINN hybrid neural network model; S4: Train the MLP-PINN hybrid neural network model based on the flow field data and operating condition characteristic parameters; S5: Predict the flow field characteristics of the quasi-rectangular section of the elbow-shaped intake channel of the pump station under unknown operating conditions based on the trained MLP-PINN hybrid neural network model.

[0006] Further, the obtaining of the flow field data of the quasi-rectangular section of the elbow-shaped intake channel under different operating conditions includes: Use particle image velocimetry experiments to obtain the boundary condition parameters of the quasi-rectangular section of the elbow inlet channel under different working conditions; According to the geometric shape of the elbow inlet channel, obtain the internal point position parameters of the uniformly distributed nodes inside the quasi-rectangular section of the elbow inlet channel.

[0007] Furthermore, the MLP-PINN hybrid neural network model includes a feedforward neural network, a connection layer, and a physics-informed neural network; Both the feedforward neural network and the physics-informed neural network include an input layer, a hidden layer, and an output layer; the connection layer connects the feedforward neural network and the physics-informed neural network by connecting the output layer of the feedforward neural network and the input layer of the physics-informed neural network.

[0008] Furthermore, training the MLP-PINN hybrid neural network model includes: A1: Normalize the obtained working condition characteristic parameters, boundary condition parameters, and internal point position parameters; A2: Input the normalized working condition characteristic parameters and boundary condition parameters into the feedforward neural network in the MLP-PINN hybrid neural network model for training to obtain the boundary conditions corresponding to the boundary of the elbow inlet channel; A3: Input the boundary conditions obtained by the feedforward neural network and the normalized internal point position parameters into the physics-informed neural network in the MLP-PINN hybrid neural network model for training to obtain the flow field data of the quasi-rectangular section of the elbow inlet channel.

[0009] Furthermore, the boundary conditions include the boundary conditions of the inlet surface and the outlet surface of the quasi-rectangular section of the elbow inlet channel.

[0010] Furthermore, the working condition characteristic parameters include the opening and closing state of the unit, the rotational speed, the water level in the intake pool, and the degree of obstruction of the trash rack.

[0011] Furthermore, the specific steps of S5 are as follows: S501: Input the normalized working condition characteristic parameters into the trained MLP-PINN hybrid neural network model, and call the feedforward neural network to predict the boundary conditions; S502: Based on the predicted boundary conditions and the point position data obtained according to the geometric shape of the elbow inlet channel, call the physics-informed neural network to predict the flow field information, and obtain the flow field characteristics of the quasi-rectangular section of the elbow inlet channel of the pump station under the corresponding working conditions.

[0012] Furthermore, the key parameters of the feedforward neural network are optimized using a hyperparameter optimization method based on Bayesian optimization to obtain the optimal parameters; The loss function of the physical information neural network is composed of a physical term and a data term, and the physical term is calculated by considering the three-dimensional NS equation of incompressible flow.

[0013] Furthermore, the key parameters of the feedforward neural network are optimized by a hyperparameter optimization method based on Bayesian optimization to obtain the optimal parameters, including: B1: Establish a learning function and confirm the sample distribution of the number of hidden layers, the number of nodes in the multi-layer perceptron, the dropout rate, the learning rate, and the optimal activation function; B2: Use Bayesian probability to determine the optimal combination of hyperparameters.

[0014] Furthermore, the physical term is calculated by considering the three-dimensional NS equation of incompressible flow, including: C1: Substitute the predicted flow field data into the partial differential equation and calculate the residual of the partial differential equation; C2: Solve the average residual of the residual as the physical term; Among them, the calculation formula of the three-dimensional NS equation is:

[0015] In the formula, is the density of water, , and are the three components of the velocity in the , and directions respectively, is the time, is the dynamic viscosity coefficient, is the pressure, is the partial derivative operator.

[0016] The beneficial effects of this application are: A flow field prediction method for a quasi-rectangular section of a pumping station channel based on a hybrid neural network provided by this application. This method conducts particle image velocimetry experiments on a model pumping station with the actual scaling ratio of the existing working conditions of the pumping station to obtain the flow field data and working condition characteristic parameters of the quasi-rectangular section of the elbow inlet channel under different working conditions, and predicts the flow field characteristics of unknown working conditions based on the trained MLP-PINN hybrid neural network model. This method predicts the flow field characteristics of other working conditions based on the experimental data of some working conditions, making up for the defects of incomplete coverage of experimental working conditions and high costs, providing more comprehensive monitoring data for the monitoring of the pumping station, and enhancing the safety of the pumping station. Description of the Drawings

[0017] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other embodiments can also be obtained based on these drawings.

[0018] Figure 1 It is a method flow chart of a flow field prediction method for a quasi-rectangular section of a pumping station channel based on a hybrid neural network provided by an embodiment of the present application.

[0019] Figure 2 It is a schematic diagram of a scaled model pumping station provided by an embodiment of the present application.

[0020] Figure 3 It is a schematic diagram of a particle image velocimetry experiment for a quasi-rectangular section of an elbow inlet channel of a pumping station provided by an embodiment of the present application.

[0021] Figure 4 It is a high-speed photography sample diagram of a quasi-rectangular section of an elbow inlet channel of a pumping station provided by an embodiment of the present application.

[0022] Figure 5 It is a cloud diagram of the flow field characteristics of a quasi-rectangular section of an elbow inlet channel of a pumping station provided by an embodiment of the present application. Specific embodiments

[0023] The following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, rather than all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art based on the present application belong to the scope of protection of the present application.

[0024] The embodiment of the present application provides a flow field prediction method for a quasi-rectangular section of a pumping station channel based on a hybrid neural network. This method can be seen in Figure 1 , Figure 1 As shown in the method flow chart of a flow field prediction method for a quasi-rectangular section of a pumping station channel based on a hybrid neural network provided by an embodiment of the present application, it includes: S1: Construct a scaled model pumping station according to the actual pumping station.

[0025] S2: Based on the model pumping station, conduct a particle image velocimetry experiment and obtain the flow field data and operating condition characteristic parameters of the quasi-rectangular section of the elbow inlet channel under different operating conditions.

[0026] Further, the obtaining of the flow field data of the quasi-rectangular section of the elbow inlet channel under different operating conditions includes: The boundary condition parameters of the quasi-rectangular section of the elbow-shaped intake channel under different working conditions are obtained by using particle image velocimetry experiments; According to the geometric shape of the elbow-shaped intake channel, the internal point position parameters of the uniformly distributed nodes inside the quasi-rectangular section of the elbow-shaped intake channel are obtained.

[0027] In an embodiment of the present application, taking an example of a flow field characteristic prediction model of the elbow-shaped intake channel of an axial-flow model pumping station, with the axial-flow pumping station operating in parallel with multiple units as the target object, the flow field characteristics of its elbow-shaped intake channel affect the stable operation and flow measurement of the pump. This model pumping station is a 1:20 scaled-down model of the actual pumping station, as Figure 2 shown. Figure 2 It is a schematic diagram of a scaled model pumping station provided by an embodiment of the present application. Four vertical axial-flow pumps operate in parallel. Taking the No. 4 pump unit as the application object, the rated speed of the unit is n = 1450 r / min, and the diameter of the pump impeller is D = 157.5 mm. The intake pool and the elbow-shaped intake channel of the model test bench are processed from transparent materials. The water level of the intake pool can be directly measured by a measuring ruler, and the degree of shielding of the guard grid is manually controlled. Particle Image Velocimetry (PIV) experiments are carried out on the model test bench, as Figure 3 shown. Figure 3 It is a schematic diagram of the particle image velocimetry experiment for the quasi-rectangular section of the elbow-shaped intake channel of the pumping station provided by an embodiment of the present application. A continuous laser is used to illuminate the tracer particles (5 - 20 microns), and the camera frame rate is adjusted to 500 hz. The purpose is to obtain the average flow field of the elbow-shaped channel over a period of time, so it is necessary to ensure that a series of captured images contain sufficient effective information. Therefore, a synchronizer is used to control the camera to continuously collect three images every second, and the total number of single samples is 1800 frames, corresponding to a duration of about 10 minutes.

[0028] When shooting the Z plane under different XY coordinates and the Y plane under different XZ coordinates, it is necessary to ensure that the laser moves horizontally in parallel and the camera position remains unchanged. The clarity is adjusted by adjusting the camera focal length, so that the XY plane and the XZ plane are respectively parallel to each other.

[0029] By controlling the water level of the intake pool, the degree of shielding of the guard grid, and the opening and closing of other units, the working conditions of the experimental unit are changed. The elbow-shaped intake channel part is photographed by a high-speed camera, and the shooting results of the PIV tracer particles are as Figure 4 shown. Figure 4This is a high-speed photography sample diagram of the quasi-rectangular section of the elbow inlet channel of the pump station provided by the embodiments of this application. Velocity vector data of the quasi-rectangular section of the elbow inlet channel is obtained through multi-frame data processing, and pressure data of evenly distributed nodes of the elbow inlet channel under different working conditions is calculated using the Navier-Stokes equation. The PIV experiment has a high measurement cost and cannot cover all operating conditions, and the flow field characteristics of the elbow inlet channel under unknown conditions are predicted based on the existing working condition data.

[0030] Examples of the characteristic parameters of the working conditions are shown in Table 1.

[0031] Table 1 Sample table of characteristic parameters of working conditions

[0032] S3: Construct an MLP-PINN hybrid neural network model.

[0033] Furthermore, the MLP-PINN hybrid neural network model includes a feedforward neural network, a connection layer, and a physics-informed neural network; Both the feedforward neural network and the physics-informed neural network include an input layer, a hidden layer, and an output layer; the connection layer connects the feedforward neural network and the physics-informed neural network by connecting the output layer of the feedforward neural network and the input layer of the physics-informed neural network.

[0034] In an embodiment of this application, a hybrid neural network model is built based on the Pytorch framework. The model includes multiple parts such as data input, data processing, the feedforward neural network MLP for predicting the boundary conditions of the inlet channel, the physics-informed neural network PINN for predicting the flow field of the inlet channel, and the output of the flow field data.

[0035] The MLP neural network part is a feedforward neural network MLP containing a multi-layer perceptron, which is used to predict the boundary conditions of the elbow inlet channel. The input is the characteristic parameters of different working conditions, and the output is an intermediate transition factor, that is, the boundary conditions of the elbow inlet channel, including the pressure and velocity of the inlet surface and the outlet surface. The key parameters of this neural network are selected using the method of hyperparameter optimization based on Bayesian optimization, including the number of hidden layers of the neural network, the number of nodes in the hidden layer, the dropout rate, the learning rate, and the activation function.

[0036] In an embodiment of this application, the key parameters of the MLP neural network are determined by establishing a learning function, specifying the sample distribution of the number of hidden layers, the number of nodes in the multi-layer perceptron, the dropout rate, the learning rate, and the optimal activation function, and determining the most promising combination of hyperparameters through Bayesian probability and iteratively adjusting the search. The key parameters of the MLP neural network are shown in Table 2.

[0037] Table 2 Table of key parameters of the MLP neural network

[0038] The Physics-Informed Neural Network (PINN) is a neural network that incorporates physical laws into the neural network training process, enabling the training of the neural network to follow physical principles. Specifically, physical equations (such as partial differential equations, PDEs) are embedded as soft constraints into the loss function of the neural network, thereby achieving efficient solution and prediction of complex physical systems. The loss function consists of a physical term and a data term. The physical term is calculated based on the three-dimensional Navier-Stokes (NS) equations considering incompressible flow. Specifically, the predicted flow field data (velocity and pressure) are substituted into the partial differential equations to calculate the residuals of the equations. The average value of the residuals of the three partial differential equations is used as the physical term residual of the PINN network, which together with the data term residual constitutes the loss function of the neural network, with the physical term and the data term having the same influence. The calculation formula for the three-dimensional NS equations is as follows:

[0039] In the formula, is the density of water, , and are the three components of the velocity in the , and directions respectively, is the time, is the dynamic viscosity coefficient, is the pressure, is the partial derivative operator.

[0040] There is a connection layer between the MLP and the PINN, which together with the output layer parameters of the MLP and the point coordinate parameters form the input layer of the PINN.

[0041] S4: Train the MLP-PINN hybrid neural network model based on the flow field data and the working condition characteristic parameters.

[0042] Furthermore, the training of the MLP-PINN hybrid neural network model includes: A1: Normalize the obtained working condition characteristic parameters, boundary condition parameters, and internal point parameters.

[0043] In an embodiment of the present application, the opening and closing state of the unit, the rotational speed, the water level in the intake sump, and the degree of blockage of the trash rack are the key factors affecting the flow field boundary conditions of the intake channel. These are used as the working condition characteristic parameters, where the opening and closing state of the unit is used as the label data, with 0 corresponding to the closed state and 1 corresponding to the open state, and the degree of blockage of the trash rack is expressed as a percentage. The boundary condition parameters can be the point data at the inlet and outlet boundaries x , y , z . The internal point parameters can be other point data inside the flow channelx , y , z .

[0044] A2: Input the normalized operating condition characteristic parameters and boundary condition parameters into the feedforward neural network in the MLP-PINN hybrid neural network model for training to obtain the boundary conditions corresponding to the boundary of the elbow inlet channel.

[0045] In an embodiment of the present application, the boundary conditions of the inlet channel flow field, as intermediate transition factors, include the boundary conditions of the inlet surface and the outlet surface, and can be each set of point data at the inlet and outlet boundaries x , y , z corresponding u , v , w , p , where u , v and w are the velocity vectors in three directions, p is the pressure.

[0046] A3: Input the boundary conditions obtained by the feedforward neural network and the normalized internal point parameters into the physics-informed neural network in the MLP-PINN hybrid neural network model for training to obtain the flow field data of the quasi-rectangular section of the elbow inlet channel of the pumping station.

[0047] In an embodiment of the present application, the flow field data of the quasi-rectangular section of the elbow inlet channel of the pumping station can be the velocities in three directions at each node u , v , w and the pressure p .

[0048] S5: Predict the flow field characteristics of the quasi-rectangular section of the elbow inlet channel of the pumping station under unknown operating conditions based on the trained MLP-PINN hybrid neural network model.

[0049] Furthermore, the specific steps of S5 include: S501: Input the normalized operating condition characteristic parameters into the trained MLP-PINN hybrid neural network model, and call the feedforward neural network to predict the boundary conditions; S502: Based on the predicted boundary conditions and the point data obtained according to the geometric shape of the elbow inlet channel, call the physics-informed neural network to predict the flow field information, and obtain the flow field characteristics of the quasi-rectangular section of the elbow inlet channel of the pumping station under the corresponding operating conditions.

[0050] In an embodiment of the present application, operating condition characteristic parameters are obtained, such as the opening and closing states of unit 1, unit 2, and unit 3, rotational speed, water level in the intake sump, degree of obstruction of the screen, and relevant parameters of the boundary conditions, such as point data at the inlet and outlet boundaries. x , y , z , and the above data is normalized, and the processed data is input into the trained MLP-PINN hybrid neural network model. The MLP part is called to calculate the corresponding flow field data at the boundary. u , v , w , p , that is, each set of point data at the inlet and outlet boundaries x , y , z corresponding to u , v , w , p . The prediction results obtained by the MLP part and other point data inside the flow channel x , y , z are input into the PINN part for prediction calculation to obtain the flow field information of all point data in the intake channel. u , v , w , p . The cloud diagram of the flow field characteristics of the quasi-rectangular section of the elbow intake channel of the pumping station predicted by the hybrid neural network is as Figure 5 shown. The figure includes the cloud diagrams of the two components of the velocity vector. The left is the velocity component u cloud diagram of the quasi-rectangular section, and the right is the velocity component v cloud diagram of the quasi-rectangular section. It can be seen that the distribution law of the velocity satisfies the pipeline velocity distribution law, verifying the accuracy of the prediction by the hybrid neural network.

[0051] In an embodiment of the present application, after obtaining the flow field information of all point data in the intake channel, the obtained data is de-normalized to obtain the real velocity and pressure, which is convenient for further flow field analysis.

[0052] A flow field prediction method for the quasi-rectangular section of a pump station flow channel based on a hybrid neural network. This method conducts particle image velocimetry experiments on a model pump station with the actual scaling ratio of the existing pump station under existing working conditions, obtains the flow field data and working condition characteristic parameters of the quasi-rectangular section of the elbow inlet flow channel under different working conditions, and predicts the flow field characteristics of unknown working conditions based on the MLP-PINN hybrid neural network model. This method predicts the flow field characteristics of other working conditions based on the experimental data of some working conditions, makes up for the defects of incomplete coverage of experimental working conditions and high costs, provides more comprehensive monitoring data for pump station monitoring, and enhances the safety of the pump station.

[0053] It should be noted that those of ordinary skill in the art will realize that the embodiments described herein are for helping readers understand the principles of the present application, and it should be understood that the protection scope of the present application is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific deformations and combinations that do not depart from the essence of the present application based on the technical revelations disclosed in the present application, and these deformations and combinations are still within the protection scope of the present application.

Claims

1. A flow field prediction method for a quasi-rectangular cross-section of a pumping station flow channel based on a hybrid neural network, characterized in that, Including: S1: Construct a scaled model pump station according to the actual pump station; S2: Based on the model pump station, conduct particle image velocimetry experiments, and obtain the flow field data and operating condition characteristic parameters of the quasi-rectangular section of the elbow inlet channel under different operating conditions; S3: Construct an MLP-PINN hybrid neural network model; S4: Train the MLP-PINN hybrid neural network model based on the flow field data and operating condition characteristic parameters; S5: Based on the trained MLP-PINN hybrid neural network model, predict the flow field characteristics of the quasi-rectangular section of the elbow inlet channel of the pump station under unknown operating conditions.

2. The flow field prediction method for the quasi-rectangular section of the pump station flow channel based on the hybrid neural network according to claim 1, wherein, The obtaining of the flow field data of the quasi-rectangular section of the elbow inlet channel under different operating conditions includes: Using particle image velocimetry experiments to obtain the boundary condition parameters of the quasi-rectangular section of the elbow inlet channel under different operating conditions; According to the geometric shape of the elbow inlet channel, obtain the internal point position parameters of the uniformly distributed nodes inside the quasi-rectangular section of the elbow inlet channel.

3. The flow field prediction method for the quasi-rectangular cross-section of the pump station flow channel based on the hybrid neural network according to claim 1, characterized in that, The MLP-PINN hybrid neural network model includes a feedforward neural network, a connection layer, and a physics-informed neural network; Both the feedforward neural network and the physics-informed neural network include an input layer, a hidden layer, and an output layer; the connection layer connects the feedforward neural network and the physics-informed neural network by connecting the output layer of the feedforward neural network and the input layer of the physics-informed neural network.

4. The flow field prediction method for the quasi-rectangular section of the pump station flow channel based on the hybrid neural network according to claim 2, characterized in that, The training of the MLP-PINN hybrid neural network model includes: A1: Normalize the obtained operating condition characteristic parameters, boundary condition parameters, and internal point position parameters; A2: Input the normalized operating condition characteristic parameters and boundary condition parameters into the feedforward neural network in the MLP-PINN hybrid neural network model for training to obtain the boundary conditions corresponding to the boundary of the elbow inlet channel; A3: Input the boundary conditions obtained by the feedforward neural network and the normalized internal point position parameters into the physics-informed neural network in the MLP-PINN hybrid neural network model for training to obtain the flow field data of the quasi-rectangular section of the elbow inlet channel.

5. The flow field prediction method for the quasi-rectangular section of the pump station flow channel based on the hybrid neural network according to claim 4, characterized in that, The boundary conditions include the boundary conditions of the inlet surface and the outlet surface of the quasi-rectangular section of the elbow inlet channel.

6. The flow field prediction method for the quasi-rectangular section of the pump station flow channel based on the hybrid neural network according to claim 4, characterized in that, The operating condition characteristic parameters include the opening and closing state of the unit, the rotational speed, the water level of the intake sump, and the degree of blockage of the trash rack.

7. The flow field prediction method for the quasi-rectangular cross-section of the pump station flow channel based on the hybrid neural network according to claim 1, wherein, The S5 specifically includes: S501: Input the normalized operating condition characteristic parameters into the trained MLP-PINN hybrid neural network model, and call the feedforward neural network to predict the boundary conditions; S502: Based on the predicted boundary conditions and the point position data obtained according to the geometric shape of the elbow inlet channel, call the physics-informed neural network to predict the flow field information, and obtain the flow field characteristics of the quasi-rectangular section of the elbow inlet channel of the pump station under the corresponding operating conditions.

8. The flow field prediction method for the quasi-rectangular cross-section of the pump station flow channel based on the hybrid neural network according to claim 3, characterized in that The key parameters of the feedforward neural network are optimized by a hyperparameter optimization method based on Bayesian optimization to obtain the optimal parameters; The loss function of the physics-informed neural network is composed of a physics term and a data term, and the physics term is calculated by considering the three-dimensional NS equation of incompressible flow.

9. The flow field prediction method for the quasi-rectangular section of the pump station flow channel based on the hybrid neural network according to claim 8, characterized in that, The key parameters of the feedforward neural network are optimized by a hyperparameter optimization method based on Bayesian optimization to obtain the optimal parameters, including: B1: Establish a learning function and confirm the sample distribution of the number of hidden layers, the number of nodes in the multi-layer perceptron, the dropout rate, the learning rate, and the optimal activation function; B2: Use Bayesian probability to determine the optimal combination of hyperparameters.

10. The flow field prediction method for the quasi-rectangular section of the pump station flow channel based on the hybrid neural network according to claim 8, characterized in that, The physical terms are calculated by considering the three-dimensional NS equations for incompressible flow, including: C1: Substitute the predicted flow field data into the partial differential equation and calculate the residual of the partial differential equation; C2: Solve the average residual of the residual as the physical term; Among them, the calculation formula of the three-dimensional NS equation is: In the formula, is the density of water, , and are the three components of the velocity in the , and directions respectively, is the time, is the dynamic viscosity coefficient, is the pressure, is the partial derivative operator.

Images