A physics and performance collaborative prediction method for pumps and hydraulic turbines

Abstract

The present invention discloses a method for collaborative prediction of physical fields and performance applicable to pumps and hydraulic turbines, comprising the steps of: 1) obtaining and encoding geometric parameters of pump and hydraulic turbine models; 2) respectively obtaining operating parameters of pump and hydraulic turbine modes and obtaining numerical simulation results; 3) normalizing input and output parameters of pump and hydraulic turbine; 4) dividing training set and verification set for pump and hydraulic turbine model data; 5) respectively constructing collaborative prediction point cloud convolution networks of two groups of physical fields and performance of pump and hydraulic turbine; 6) training two groups of point cloud convolution neural networks of pump and hydraulic turbine; 7) post-processing performance prediction information. The present invention realizes direct mapping from relevant performance parameters under pump mode to relevant performance parameters under hydraulic turbine mode, solves the shortcomings of large amount of calculation data and long calculation time of traditional numerical simulation methods, and has the advantages of easy data modification and timely acquisition of required performance parameters.

Description

Technical Field

[0001] The invention belongs to the technical field of pumps and hydraulic turbines, and in particular relates to a physical field and performance collaborative prediction method applicable to pumps and hydraulic turbines. Background Art

[0002] At present, there is often a large amount of clean, green, pollution-free, and surplus pressure energy in the industrial production and transmission processes in the fields of petroleum, chemical industry, etc. This pressure energy can be fully utilized by pump reversal to make hydraulic turbines. At present, most of the research on pump reversal to make hydraulic turbines uses numerical simulation methods. This method obtains the flow field of the pump when it is reversed under different working conditions by performing three-dimensional steady flow field numerical simulation calculations on the internal flow field of the pump when it is reversed, thereby obtaining the performance of the pump under different working conditions. The use of three-dimensional numerical simulation calculations requires a lot of resources and time. When the relevant working parameters and geometric parameters change, the corresponding performance parameters cannot be obtained in time, resulting in a great waste of resources and time lag.

[0003] In recent years, with the development of science and technology, the computing power of computers has been significantly improved. The progress of artificial intelligence has brought many new solutions to traditional problems, and deep learning methods have been promoted and disseminated in various engineering fields. Summary of the invention

[0004] The purpose of the present invention is to provide a physical field and performance collaborative prediction method suitable for pumps and hydraulic turbines. The method encodes the geometric parameters of the blades in the pump, and inputs the encoded data together with the operating parameters of the pump and the hydraulic turbine into two groups of networks, each of which contains two point cloud convolutional neural networks. In the first group, the first P_Net1 network predicts the pump physical field by the operating parameters and geometric parameters of the pump, and the second P_Net2 network predicts the pump performance by the pump physical field; in the second group, the first T_Net1 network predicts the hydraulic turbine physical field by the operating parameters and geometric parameters of the hydraulic turbine, and the second T_Net2 network predicts the hydraulic turbine performance by the hydraulic turbine physical field. Finally, by fitting the two groups of data, a mapping curve from the pump performance under variable working conditions to the hydraulic turbine performance is obtained, thereby completing the task of multi-physical field and performance collaborative prediction of the hydraulic turbine under variable working conditions.

[0005] The present invention is achieved by adopting the following technical solutions:

[0006] A physical field and performance collaborative prediction method applicable to pumps and hydraulic turbines includes the following steps:

[0007] 1) Obtain the geometric parameters of the pump and hydraulic turbine model and encode them

[0008] The obtained geometric parameters of the blades are the installation angle α of the control blades, the number of blades a, and the coordinates (x, y, z) of the B points of the Bezier curve of the control blades. Each pump corresponds to a blade. Since the same blades are used in the pump and hydraulic turbine modes, the initial geometric parameter sets in both modes are obtained as Modeling is performed based on the geometric parameters of the I pumps, and for each grid model, its grid node information G is obtained. i,p,q , the grid node information G i,p,q The initial geometric parameters of the blade are set Integrate to obtain the geometric parameter set A i,j , and the geometric parameters are encoded by an encoder-decoder combination. The encoder Encoder encodes the geometric parameter set A of the blade i,j Mapped into a fixed-length vector, the decoder decodes the encoded vector into a geometric parameter set A i,j , the vector combination of I leaves obtained by the encoder constitutes the indirect geometric parameter set A of the leaf i,f ; Wherein, i = 1, 2, 3, ..., I, represents different pumps, I is the total number of pumps, j = 1, 2, 3, ..., J, represents different geometric parameters, J is the total number of geometric parameters, p = 1, 2, 3, ..., P, represents different grid nodes, P is the total number of grid nodes, q is the coordinate of each grid node, including the coordinates in the x-, y- and z-directions, and if it is a two-dimensional model, it includes the coordinates in the x- and y-directions, f = 1, 2, 3, ..., F, represents different indirect geometric parameter variables, F is the total number of indirect geometric parameter variables;

[0009] 2) Obtain the operating parameters of the pump and hydraulic turbine modes respectively and obtain the numerical simulation results

[0010] The operating parameters of the pump and hydraulic turbine include the mechanical speed ω when the pump and hydraulic turbine are running in two modes. R , pump and hydraulic turbine inlet pressure P in and inlet temperature T in ; For each pump, in the two modes, the Latin hypercube sampling method is used to randomly sample D working conditions in their respective parameter spaces. The working condition parameter set under the pump mode is B i,d , the operating parameter set under the hydraulic turbine mode is C i,d Each pump model is subjected to CFD calculations at D operating points in two modes to obtain a three-dimensional physical field data set for each blade in the pump mode. Combined with 3D physical field data in hydraulic turbine mode The performance parameter set of different pumps under pump mode is calculated based on the physical field parameters and performance parameter set in hydraulic turbine mode Among them, d = 1, 2, 3, ..., D, represents different operating parameters, D is the total number of operating points that need to be calculated in each mode, n = 1, 2, 3, ..., N, represents different physical fields, including velocity, temperature, pressure and vorticity, N is the total number of recorded flow field parameters, m = 1, 2, respectively represent efficiency and head;

[0011] 3) Normalization of pump and hydraulic turbine input and output parameters

[0012] The simplified set and other data are normalized. The input data set that needs to be normalized is {A i,f ,B i,d ,C i,d}, the physical field set is The performance parameter set is The normalized input data set is The physics set is The performance parameter set is

[0013] 4) Divide the pump and hydraulic turbine model data into training and validation sets

[0014] For neural networks such as point clouds, the input data set is The input data set Randomly shuffle and divide the input set into a training set in a ratio of 2:2:1 Validation set and test set Likewise, the physics collection is divided into and The performance parameter set is also divided into and

[0015] 5) Construct point cloud convolutional networks for collaborative prediction of the physical fields and performance of pumps and hydraulic turbines respectively

[0016] Construct two groups of networks: pump physics and performance prediction and hydraulic turbine physics and performance prediction. Each group contains two point cloud convolutional neural networks. For the first group of pump physics and performance prediction networks, the first P_Net1 point cloud convolutional neural network is based on the input set in the pump mode. Predict the physical fields in pump mode Realize the mapping from design variables, i.e., geometric parameters and operating parameters, to the pump physical field in the pump mode. The second P_Net2 point cloud convolutional neural network input is the pump physical field set predicted by the P_Net1 network. Predict the performance parameters set in pump mode Realize the mapping from pump physical field to pump performance parameters; for the second set of hydraulic turbine physical field and performance prediction network, the first T_Net1 point cloud convolutional neural network is based on the input set under the hydraulic turbine mode Predict the physical fields in the hydraulic turbine mode Realize the mapping from design variables to hydraulic turbine physical fields in hydraulic turbine mode. The second T_Net2 point cloud convolutional neural network input is the physical field predicted by the T_Net1 network Predict the performance parameters set under hydraulic turbine mode Realize the mapping from hydraulic turbine physical field to hydraulic turbine performance parameters;

[0017] 6) Training two sets of point cloud convolutional neural networks for pumps and hydraulic turbines

[0018] 7) Post-processing of performance prediction information

[0019] According to the output set of two parts and A direct connection between the performance parameters under pump mode and the performance parameters under hydraulic turbine mode is established, and a direct mapping between the efficiency under pump mode and the efficiency under hydraulic turbine mode is achieved. At the same time, the relevant physical field analysis of the pump and hydraulic turbine is obtained for further analysis.

[0020] A further improvement of the present invention is that it also includes the following steps:

[0021] 8) Algorithm maintenance

[0022] In actual application, if the input information needs to increase or decrease, the encoder and decoder are used to map the original input information into a vector of fixed size as a new input. The four sets of point cloud convolutional networks for pump and hydraulic turbine physical fields and performance prediction that have been trained are used as pre-trained models, and re-training is performed on this basis.

[0023] A further improvement of the present invention is that in step 1), both the encoder Encoder and the decoder Decoder use a fully connected layer to construct a network, and the encoder Encoder constructs a coding mapping function to realize the geometric parameter set A i,j The decoder constructs a decoding mapping function to realize the encoding mapping from the fixed-length vector to the geometric parameter set A. i,j The decoding mapping of the process is the loss function of the geometric parameter set A i,j The predicted geometric parameter result after encoding and decoding is different from the original true geometric parameter set A i,j The difference between:

[0024]

[0025] A further improvement of the present invention is that in step 3), the normalization operation is a Sigmoid function conversion, and the specific operation is as follows:

[0026]

[0027] Among them, the expression of the Sigmoid function is as follows:

[0028]

[0029] The normalization operation of the remaining data is similar:

[0030]

[0031]

[0032]

[0033]

[0034]

[0035]

[0036] A further improvement of the present invention is that in step 4), a point cloud convolutional network is used for collaborative prediction of the physical fields and performance of pumps and hydraulic turbines.

[0037] A further improvement of the present invention is that the point cloud convolutional network adopts PointNet, PointNet++ or Kd-Network.

[0038] A further improvement of the present invention is that in step 5), for the two modes of pump and hydraulic turbine, the two point cloud convolutional neural networks in each mode include two parts of loss functions, one part is the loss function F_loss of physical field prediction, which is used to reduce the gap between the predicted physical field and the real physical field; the other part is the loss function ψ_loss of performance prediction, which is used to reduce the gap between the predicted performance and the real performance;

[0039]

[0040]

[0041]

[0042]

[0043] Among them, w qThe weight of the square difference between the predicted physical field data and the real physical field data at each network node, These are the predicted physical field data generated by the two physical field prediction point cloud convolutional neural networks P_Net1 and T_Net1. are the real physical field data of the pump and hydraulic turbine respectively, These are the predicted performance data generated by the two performance prediction point cloud convolutional neural networks P_Net2 and T_Net2. These are the actual performance data of the pump and hydraulic turbine respectively.

[0044] A further improvement of the present invention is that in step 6), during the network training process, the optimizer is first set to Adam, the initial learning rate is set to 0.01, and training is performed for 50 steps; then the optimizer is set to SGD, and then the learning rate is reduced to 1 / 10 of the original value every 100 steps of training, and the minimum learning rate is 0.0001.

[0045] The present invention has at least the following beneficial technical effects:

[0046] The present invention provides a method for collaborative prediction of physical fields and performance applicable to pumps and hydraulic turbines. The method encodes the geometric parameters of the blades in the pump, and inputs the encoded data together with the operating parameters of the pump and hydraulic turbine into two groups of networks, each of which contains two point cloud convolutional neural networks. The first P_Net1 network in the first group predicts the pump physical field by the operating parameters and geometric parameters of the pump, and the second P_Net2 network predicts the pump performance by the pump physical field; the first T_Net1 network in the second group predicts the hydraulic turbine physical field by the operating parameters and geometric parameters of the hydraulic turbine, and the second T_Net2 network predicts the hydraulic turbine performance by the hydraulic turbine physical field. Finally, the mapping curve of the pump performance to the hydraulic turbine performance under variable working conditions is obtained by fitting the two groups of data. The establishment of this method can realize the direct mapping from the relevant performance parameters under the pump mode to the relevant performance parameters under the hydraulic turbine mode, which solves the shortcomings of traditional numerical simulation methods such as large amount of calculation data and long calculation time. It has the advantages of easy data change and timely acquisition of required performance parameters. It can replace traditional numerical simulation calculations and has the advantages of low cost and high accuracy compared to experimental methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a flow chart of the multi-physical field and performance collaborative prediction method of a variable operating condition hydraulic turbine based on data-driven according to the present invention;

[0048] Figure 2 It is an explanatory diagram of the data processing process of the present invention;

[0049] Figure 3A schematic diagram of a flow field prediction point cloud convolutional neural network constructed for the present invention;

[0050] Figure 4 Schematic diagram of the performance prediction point cloud convolutional neural network constructed for the present invention. DETAILED DESCRIPTION

[0051] The present invention is further described in detail below in combination with the embodiments according to the content of the invention. The following is an application of the present invention, but is not limited thereto. The implementers may modify the parameters according to the specific circumstances. The specific implementation method is as follows: Figure 1 As shown in Figure 2 shown.

[0052] like Figure 1 As shown, the present invention provides a physical field and performance collaborative prediction method applicable to pumps and hydraulic turbines, comprising the following steps:

[0053] The first step is to obtain the geometric parameters of the pump and hydraulic turbine models and encode them;

[0054] Get the geometric parameters of the light sheet, such as the installation angle α of the control blade, the number of blades a, and the coordinates (x, y, z) of the B points of the Bezier curve of the control blade. Each pump corresponds to a blade. Since the same blades are used in the pump and hydraulic turbine modes, the initial geometric parameter sets in both modes are obtained as Modeling is performed based on the geometric parameters of the I pumps, and for each grid model, its grid node information G is obtained. i,p,q , the grid node information G i,p,q The initial geometric parameters of the blade are set Integrate to obtain the geometric parameter set A i,j In order to prevent the dimensional disaster caused by the large number of geometric parameters of the blade, an encoder-decoder combination is used to encode the geometric parameters. The encoder can encode the geometric parameter set A of the blade i,j Mapped into a fixed-length vector, the decoder can decode the encoded vector into a geometric parameter set A i,j , the loss function of this process is the geometric parameter set A i,j The predicted geometric parameter result after encoding and decoding is different from the original true geometric parameter set A i,j The difference between:

[0055]

[0056] The vector combination of I leaves obtained by the encoder constitutes the indirect geometric parameter set A of the leaves i,f. Among them, i=1,2,3,...,I, represents different pumps, I is the total number of pumps, j=1,2,3,...,J, represents different geometric parameters, J is the total number of geometric parameters, p=1,2,3,...,P, represents different grid nodes, P is the total number of grid nodes, q is the coordinate of each grid node, including the coordinates in the x-direction, y-direction and z-direction. If it is a two-dimensional model, it only contains the coordinates in the x-direction and y-direction. f=1,2,3,...,F, represents different indirect geometric parameter variables,

[0057] The second step is to obtain the operating parameters of the pump and hydraulic turbine modes and obtain the numerical simulation results;

[0058] Get the mechanical speed ω when the pump and hydraulic turbine are running in two modes R , pump and hydraulic turbine inlet pressure P in and inlet temperature T in For each pump, in the two modes, the Latin hypercube sampling method is used to randomly sample D operating conditions in the respective parameter space. The operating condition parameter set under the pump mode is B i,d , the operating parameter set under the hydraulic turbine mode is C i,d Each pump model is subjected to CFD calculations at D operating points in two modes to obtain a set of physical field data for each blade in the pump mode. Combined with the physical field data in the hydraulic turbine model According to the physical field parameters, the performance parameter set of different pumps in the pump mode can be calculated and performance parameter set in hydraulic turbine mode Among them, d = 1, 2, 3, ..., D, represents different operating parameters, D is the total number of operating points that need to be calculated in each mode, n = 1, 2, 3, ..., N, represents different physical fields, including velocity, temperature, pressure and vorticity, etc., N is the total number of recorded flow field parameters, and m = 1, 2, respectively represent efficiency and head.

[0059] The third step is to normalize the input and output parameters of the pump and hydraulic turbine;

[0060] The simplified set and other data are normalized. The input data set that needs to be normalized is {A i,f ,B i,d ,C i,d}, the physical field set is The performance parameter set is The normalized input data set is The physics set is The performance parameter set is

[0061] The normalization operation is Sigmoid function conversion, the specific operation is as follows:

[0062]

[0063] Among them, the expression of the Sigmoid function is as follows:

[0064]

[0065] The normalization operation of the remaining data is similar:

[0066]

[0067]

[0068]

[0069]

[0070]

[0071]

[0072] The normalized input data set is The physics set is The performance parameter set is

[0073] The fourth step is to divide the pump and hydraulic turbine model data into training sets and validation sets;

[0074] For the PointNet point cloud convolutional neural network, the input data set is The input data set Randomly shuffle and divide the input set into a training set in a ratio of 2:2:1 Validation set and test set Similarly, the physics field set should be divided into and The performance parameter set is also divided into and

[0075] The fifth step is to construct collaborative prediction point cloud convolutional networks for the physical fields and performance of the pump and hydraulic turbine respectively;

[0076] Construct two groups of networks: pump physics and performance prediction and hydraulic turbine physics and performance prediction. Each group contains Figure 3 and Figure 4For the first set of pump physics and performance prediction networks, the first P_Net1 point cloud convolutional neural network is based on the input set under the pump mode. Predict the physical fields in pump mode To realize the mapping from design variables (geometric parameters and operating parameters) to the pump physical field in the pump mode, the loss function of the network is as follows:

[0077]

[0078] The second P_Net2 point cloud convolutional neural network predicts the pump physics set obtained by the P_Net1 network. Predict the performance parameters set in pump mode To achieve the mapping from the pump physical field to the pump performance parameters, the loss function of the network is as follows:

[0079]

[0080] For the second set of hydraulic turbine physical field and performance prediction networks, the first T_Net1 point cloud convolutional neural network is based on the input set under the hydraulic turbine mode. Predict the physical fields in the hydraulic turbine mode The loss function of the network is used to map the design variables (geometric parameters and operating parameters) to the physical field of the hydraulic turbine in the hydraulic turbine mode.

[0081]

[0082] The second T_Net2 point cloud convolutional neural network predicts the physical field obtained by the T_Net1 network Predict the performance parameters set under hydraulic turbine mode To achieve the mapping from the hydraulic turbine physical field to the hydraulic turbine performance parameters, the loss function of the network is as follows:

[0083]

[0084] Step 6: Train four groups of point cloud convolutional neural networks for the pump and hydraulic turbine.

[0085] In the process of training the network, the optimizer is first set to Adam, the initial learning rate is set to 0.01, and training is performed for 50 steps; then the optimizer is set to SGD, and then the learning rate is reduced to 1 / 10 of the original every 100 steps of training, and the minimum learning rate is 0.0001.

[0086] Step 7: Post-processing of performance prediction information;

[0087] According to the output set of two parts and A direct connection between the performance parameters in pump mode and those in hydraulic turbine mode is established, and a direct mapping between the efficiency in pump mode and the efficiency in hydraulic turbine mode is achieved. At the same time, the relevant physical field analysis of the pump and hydraulic turbine is obtained for further analysis.

[0088] 8) Algorithm maintenance

[0089] In actual application, if the input information needs to increase or decrease, the encoder and decoder are used to map the original input information into a vector of fixed size as a new input. The four sets of point cloud convolutional networks for pump and hydraulic turbine physical fields and performance prediction that have been trained are used as pre-trained models, and re-training is performed on this basis.

[0090] In step 2), the operating conditions are divided into Latin hypercube sampling. The points taken in the operating condition parameter space by this method are multi-dimensional layered points. This method ensures that the obtained operating condition points can cover the entire parameter space well. At the same time, this method is also closest to random sampling.

[0091] Although the present invention has been described in detail above with general descriptions and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements may be made thereto based on the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all fall within the scope of protection claimed by the present invention.

Claims

1. A physical field and performance collaborative prediction method applicable to pumps and hydraulic turbines, characterized in that: The following steps are involved: 1) Obtain the geometric parameters of the pump and hydraulic turbine model and encode them The obtained geometric parameters of the blades are the installation angle α of the control blades, the number of blades a, and the coordinates (x, y, z) of the B points of the Bezier curve of the control blades. Each pump corresponds to a blade. Since the same blades are used in the pump and hydraulic turbine modes, the initial geometric parameter sets in both modes are obtained as Modeling is performed based on the geometric parameters of the I pumps, and for each grid model, its grid node information G is obtained. i,p,q , the grid node information G i,p,q The initial geometric parameters of the blade are set Integrate to obtain the geometric parameter set A i,j , and the geometric parameters are encoded by an encoder-decoder combination. The encoder Encoder encodes the geometric parameter set A of the blade i,j Mapped into a fixed-length vector, the decoder decodes the encoded vector into a geometric parameter set A i,j , the vector combination of I leaves obtained by the encoder constitutes the indirect geometric parameter set A of the leaf i,f ; Wherein, i = 1, 2, 3, ..., I, represents different pumps, I is the total number of pumps, j = 1, 2, 3, ..., J, represents different geometric parameters, J is the total number of geometric parameters, p = 1, 2, 3, ..., P, represents different grid nodes, P is the total number of grid nodes, q is the coordinate of each grid node, including the coordinates in the x-, y- and z-directions, and if it is a two-dimensional model, it includes the coordinates in the x- and y-directions, f = 1, 2, 3, ..., F, represents different indirect geometric parameter variables, F is the total number of indirect geometric parameter variables; 2) Obtain the operating parameters of the pump and hydraulic turbine modes respectively and obtain the numerical simulation results The operating parameters of the pump and hydraulic turbine include the mechanical speed ω when the pump and hydraulic turbine are running in two modes. R , pump and hydraulic turbine inlet pressure P in and inlet temperature T in ; For each pump, in the two modes, the Latin hypercube sampling method is used to randomly sample D working conditions in their respective parameter spaces. The working condition parameter set under the pump mode is B i,d , the operating parameter set under the hydraulic turbine mode is C i,d Each pump model is subjected to CFD calculations at D operating points in two modes to obtain a three-dimensional physical field data set for each blade in the pump mode. Combined with 3D physical field data in hydraulic turbine mode The performance parameter set of different pumps under pump mode is calculated based on the physical field parameters and performance parameter set in hydraulic turbine mode Among them, d = 1, 2, 3, ..., D, represents different operating parameters, D is the total number of operating points that need to be calculated in each mode, n = 1, 2, 3, ..., N, represents different physical fields, including velocity, temperature, pressure and vorticity, N is the total number of recorded flow field parameters, m = 1, 2, respectively represent efficiency and head; 3) Normalization of pump and hydraulic turbine input and output parameters The simplified set and other data are normalized. The input data set that needs to be normalized is {A i,f ,B i,d ,C i,d }, the physical field set is The performance parameter set is The normalized input data set is The physics set is The performance parameter set is 4) Divide the pump and hydraulic turbine model data into training and validation sets For neural networks such as point clouds, the input data set is The input data set Randomly shuffle and divide the input set into a training set in a ratio of 2:2:1 Validation set and test set Likewise, the physics collection is divided into and The performance parameter set is also divided into and 5) Construct point cloud convolutional networks for collaborative prediction of the physical fields and performance of pumps and hydraulic turbines respectively Construct two groups of networks: pump physics and performance prediction and hydraulic turbine physics and performance prediction. Each group contains two point cloud convolutional neural networks. For the first group of pump physics and performance prediction networks, the first P_Net1 point cloud convolutional neural network is based on the input set in the pump mode. Predict the physical fields in pump mode Realize the mapping from design variables, i.e., geometric parameters and operating parameters, to the pump physical field in the pump mode. The second P_Net2 point cloud convolutional neural network input is the pump physical field set predicted by the P_Net1 network. Predict the performance parameters set in pump mode Realize the mapping from pump physical field to pump performance parameters; for the second set of hydraulic turbine physical field and performance prediction network, the first T_Net1 point cloud convolutional neural network is based on the input set under the hydraulic turbine mode Predict the physical fields in the hydraulic turbine mode Realize the mapping from design variables to hydraulic turbine physical fields in hydraulic turbine mode. The second T_Net2 point cloud convolutional neural network input is the physical field predicted by the T_Net1 network Predict the performance parameters set under hydraulic turbine mode Realize the mapping from hydraulic turbine physical field to hydraulic turbine performance parameters; 6) Training two sets of point cloud convolutional neural networks for pumps and hydraulic turbines 7) Post-processing of performance prediction information According to the output set of two parts and A direct connection between the performance parameters under pump mode and the performance parameters under hydraulic turbine mode is established, and a direct mapping between the efficiency under pump mode and the efficiency under hydraulic turbine mode is achieved. At the same time, the relevant physical field analysis of the pump and hydraulic turbine is obtained for further analysis.

2. A physical field and performance collaborative prediction method applicable to pumps and hydraulic turbines according to claim 1, characterized in that: The following steps are also included: 8) Algorithm maintenance In actual application, if the input information needs to increase or decrease, the encoder and decoder are used to map the original input information into a vector of fixed size as a new input. The four sets of point cloud convolutional networks for pump and hydraulic turbine physical fields and performance prediction that have been trained are used as pre-trained models, and re-training is performed on this basis.

3. A physical field and performance collaborative prediction method applicable to pumps and hydraulic turbines according to claim 1, characterized in that: In step 1), both the encoder and decoder use fully connected layers to build the network. The encoder constructs the encoding mapping function to realize the transformation from the geometric parameter set A to i,j The decoder constructs a decoding mapping function to realize the encoding mapping from the fixed-length vector to the geometric parameter set A. i,j The decoding mapping of the process is the loss function of the geometric parameter set A i,j The predicted geometric parameter result after encoding and decoding is different from the original true geometric parameter set A i,j The difference between:

4. A physical field and performance collaborative prediction method applicable to pumps and hydraulic turbines according to claim 1, characterized in that: In step 3), the normalization operation is Sigmoid function conversion, and the specific operation is as follows: Among them, the expression of the Sigmoid function is as follows: The normalization operation of the remaining data is similar:

5. A physical field and performance collaborative prediction method applicable to pumps and hydraulic turbines according to claim 1, characterized in that: In step 4), a point cloud convolutional network is used for collaborative prediction of pump and hydraulic turbine physics and performance.

6. A physical field and performance collaborative prediction method applicable to pumps and hydraulic turbines according to claim 5, characterized in that: The point cloud convolutional network uses PointNet, PointNet++ or Kd-Network.

7. A physical field and performance collaborative prediction method applicable to pumps and hydraulic turbines according to claim 1, characterized in that: In step 5), for the two modes of pump and hydraulic turbine, the two point cloud convolutional neural networks in each mode contain two parts of loss functions, one part is the loss function F_loss of physical field prediction, which is used to reduce the gap between the predicted physical field and the real physical field; the other part is the loss function ψ_loss of performance prediction, which is used to reduce the gap between the predicted performance and the real performance; Among them, w q The weight of the square difference between the predicted physical field data and the real physical field data at each network node, These are the predicted physical field data generated by the two physical field prediction point cloud convolutional neural networks P_Net1 and T_Net1. are the real physical field data of the pump and hydraulic turbine respectively, These are the predicted performance data generated by the two performance prediction point cloud convolutional neural networks P_Net2 and T_Net2. These are the actual performance data of the pump and hydraulic turbine respectively.

8. A physical field and performance collaborative prediction method applicable to pumps and hydraulic turbines according to claim 1, characterized in that: In step 6), during the network training process, the optimizer is first set to Adam, the initial learning rate is set to 0.01, and training is performed for 50 steps; then the optimizer is set to SGD, and then the learning rate is reduced to 1 / 10 of the original value every 100 steps of training, and the minimum learning rate is 0.0001.

Images