CN114692523A - Flow velocity prediction method of self-adaptive high-dimensional hydrodynamic equation based on graph convolution - Google Patents
Flow velocity prediction method of self-adaptive high-dimensional hydrodynamic equation based on graph convolution Download PDFInfo
- Publication number
- CN114692523A CN114692523A CN202210300040.2A CN202210300040A CN114692523A CN 114692523 A CN114692523 A CN 114692523A CN 202210300040 A CN202210300040 A CN 202210300040A CN 114692523 A CN114692523 A CN 114692523A
- Authority
- CN
- China
- Prior art keywords
- points
- grid
- layer
- point
- equation
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 68
- 239000012530 fluid Substances 0.000 claims abstract description 49
- 238000013528 artificial neural network Methods 0.000 claims abstract description 18
- 230000008569 process Effects 0.000 claims abstract description 11
- 238000012821 model calculation Methods 0.000 claims abstract description 4
- 239000011159 matrix material Substances 0.000 claims description 48
- 230000006870 function Effects 0.000 claims description 21
- 238000012549 training Methods 0.000 claims description 20
- 230000003044 adaptive effect Effects 0.000 claims description 18
- 235000015220 hamburgers Nutrition 0.000 claims description 10
- 230000004907 flux Effects 0.000 claims description 8
- 230000004931 aggregating effect Effects 0.000 claims description 6
- 238000003062 neural network model Methods 0.000 claims description 6
- 238000010276 construction Methods 0.000 claims description 5
- 238000005070 sampling Methods 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 3
- 239000011362 coarse particle Substances 0.000 claims 1
- 230000035939 shock Effects 0.000 claims 1
- 238000004364 calculation method Methods 0.000 description 13
- 238000004088 simulation Methods 0.000 description 9
- 239000002826 coolant Substances 0.000 description 6
- 238000013527 convolutional neural network Methods 0.000 description 4
- 238000006073 displacement reaction Methods 0.000 description 4
- 238000013507 mapping Methods 0.000 description 3
- 238000007796 conventional method Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 238000007792 addition Methods 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 229920001971 elastomer Polymers 0.000 description 1
- 239000000806 elastomer Substances 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 239000003758 nuclear fuel Substances 0.000 description 1
- 230000002035 prolonged effect Effects 0.000 description 1
- 238000001228 spectrum Methods 0.000 description 1
- 239000004575 stone Substances 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/28—Design optimisation, verification or simulation using fluid dynamics, e.g. using Navier-Stokes equations or computational fluid dynamics [CFD]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Evolutionary Computation (AREA)
- Strategic Management (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- General Engineering & Computer Science (AREA)
- Human Resources & Organizations (AREA)
- Economics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Fluid Mechanics (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Mathematical Physics (AREA)
- Pure & Applied Mathematics (AREA)
- Algebra (AREA)
- Development Economics (AREA)
- Computing Systems (AREA)
- Game Theory and Decision Science (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Entrepreneurship & Innovation (AREA)
- Marketing (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention belongs to the technical field of fluid dynamic field prediction and discloses a flow velocity prediction method of a self-adaptive high-dimensional fluid dynamic equation based on graph convolution, which comprises the steps of firstly establishing a discretization grid by utilizing a high-dimensional fluid dynamic equation space structure model of a graph convolution neural network and establishing a self-adaptive space structure relation for each grid point in the grid by utilizing the graph convolution neural network; then, based on a finite difference differential equation numerical method, through spatial structure model calculation, predicting and outputting spatial structure coefficients of each grid point; further calculating a spatial derivative; and finally, obtaining a final numerical solution of the high-dimensional hydrodynamic equation based on a physical solving process of the high-dimensional hydrodynamic equation.
Description
Technical Field
The invention belongs to the technical field of fluid dynamic field prediction, and particularly relates to a flow velocity prediction method of a self-adaptive high-dimensional fluid dynamic equation based on graph convolution.
Background
In the prediction of the fluid dynamic field, the discretization of the fluid dynamic partial differential equation is generally used for solving to obtain an approximate value. The key process of discretization of partial differential equations is to calculate prediction coefficients, and methods such as finite difference or finite volume are mainly used. The conventional method mainly faces the problems that the rough grid division can not obtain an accurate solution and the calculation speed is slow, because the conventional method depends on the size of the grid, and when the division is very fine, the calculation time is prolonged and the calculation speed is slow. Therefore, a data-driven discretization method that systematically derives the discretization of a continuous physical system using machine learning has been proposed. The method has the advantages that the convolutional neural network is used for replacing the computation of the finite difference coefficient, the rest steps are still performed by using the traditional method, the convolutional neural network is integrated into the learning process of the user embedding, and compared with a method without considering comment information, the obtained representation improves the generalization of the user embedding. However, this method still has the following problems:
first, a structured representation of the velocity node association cannot be established (structural relationship missing). Because the method only considers the influence of a single node on the coefficient, the structural relationship between the nodes is not mined. For example, the current method only considers the speed information of the nodes when calculating the coefficients by using the neural network, ignores the relationship between the nodes, and not only lacks rationality of a space mapping mechanism, but also reduces the generalization of the representation.
Secondly, the method only provides a low-dimensional space structure modeling scheme, so that the method is applied to a low-dimensional fluid dynamic equation and cannot solve the problem of a complex space structure of a high-dimensional fluid dynamic equation.
Therefore, in order to solve the above problems, the present invention provides a flow velocity prediction method based on an adaptive high-dimensional fluid dynamic equation of graph convolution.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides a flow velocity prediction method of a graph convolution-based self-adaptive high-dimensional hydrodynamic equation, wherein a spatial structure relation is established for each grid point velocity value through a convolution neural network, and then the velocity value at the next moment is obtained through prediction.
In order to solve the technical problems, the invention adopts the technical scheme that:
the flow velocity prediction method of the self-adaptive high-dimensional hydrodynamic equation based on graph convolution comprises the following steps:
s1, node x of fluidnInputting the coarse-grained speed function value into a high-dimensional fluid dynamic equation space structure model of a graph convolution neural network, and constructing an adaptive space structure relation for each grid point in the grid by establishing a discretization grid by using the graph convolution neural network;
s2, predicting and outputting the space structure coefficient of each grid point by a finite difference-based differential equation numerical method through space structure model calculation
S3, any grid point x obtained by predicting step S2nCoefficient of spatial structure ofAnd xnX of surrounding pointsn-k,…,xn+kThe coarseness values are combined, i.e. xnPoint and xnK points on the left and k points on the right are adjacent, and x is calculatednSpatial derivatives of (i.e. x)nProcessing the solution of the velocity value after the spatial evolution;
s4, physical solving process based on high-dimensional fluid dynamic equation, namely calculating flux and then calculating x by utilizing the fluxnTime derivative of (i.e. x)nAnd (4) solving the velocity value after the velocity value evolves along with time to finally obtain the final numerical solution of the high-dimensional fluid dynamic equation.
Further, when the adaptive spatial structure relationship is constructed in step S1, the feature information of each node itself is obtained, and the feature information of the node is obtained from all neighboring nodes of each node, so as to construct a graph structure relationship, including:
2) construction of graph structure relationships for one-dimensional fluid dynamic equations
The one-dimensional spatial equation for Burgers is as follows:
Known training data set has T1A time layer, each layer having M1A number of grid points, each having speed information thereon, M1Down sampling at each grid point to obtain m1A plurality of grid points; when training the model, firstly, the m of the first layer is1The speed values of the grid points are used as input, loss is obtained by calculating MAE through the true solution of the speed value of the next time layer obtained through training and the speed value of the next time layer, and when the layer is trained, the next layer is trained;
by constructing a graph structure of a discretization grid, taking spatial grid points as vertexes of the graph, designing relationship weights between every two different grid points as edges of the graph, and establishing spatial structure relationships of all the grid points; the method comprises the following specific steps: when establishing spatial structure relationship, all m of each layer1Establishing a spatial structure relationship at each grid point, wherein the velocity value of each grid point is represented by the velocity values of 2k grid points around the upper layer and the velocity value at the grid point, each grid point has velocity information, and thus the input feature matrix X is obtained from the velocity of each grid point, and the shape of the X is (m)11); since each grid point is represented by 2k points around, i.e. k points to the left and k points to the right of the point; for the first three points, i.e. P1,P2,P3The left side is less than k points, so the three points are represented by the right k points and all the points on the left side of the points, and a space structure relationship is established;
the adjacency matrix A can be obtained by establishing a spatial structure relationship, and the adjacency matrix A can be obtained by the adjacency matrixAnddegree matrix ofIs given by the formulaWherein,i is an identity matrix formed by connecting X,Substituting the speed information into a formula (2), aggregating the speed information of each grid point, predicting to obtain a finite difference coefficient, and calculating a spatial difference and a time derivative;
in addition, the propagation mode between layers of the graph convolution neural network is as follows:
wherein, H is a feature matrix of each layer, obtained from the speed of all grid points of each layer, and is X for the input layer; σ is a nonlinear activation function, which both makes full use of the velocity information of the grid points and their relationship, and enhances the correlation between the grid points.
Further, in establishing the spatial structure relationship, m for each layer1Establishing a spatial structure relationship at each grid point, wherein the velocity value of each grid point is represented by the velocity values of 6 grid points around the previous layer and the velocity value at the grid point, each grid point has velocity information, and thus the input feature matrix X is obtained from the velocity of each grid point, and the shape of the X is (m)11); since each grid point is represented by 6 surrounding points, i.e. 3 points to the left and 3 points to the right of the point, for the first three points, i.e. P1,P2,P3The left side is less than 3 points, so the three points are represented by the right 3 points and all the points on the left side of the points, so that the spatial structure relationship is established.
Further, 2) constructing a graph structure relation for a two-dimensional fluid dynamic equation:
the two-dimensional spatial equation for Burgers is as follows:
known training data set has T2A time layer, each layer having M because of two dimensions of x and y1×M1A plurality of grid points, each grid point having speed information, M1×M1Down sampling at each grid point to obtain m2×m2A plurality of grid points;
when building spatial structure relationship, m for each layer2×m2Establishing a spatial structure relationship at each grid point, wherein the speed value of each grid point is represented by the speed values of 4 grid points around the point at the upper layer and the speed value of the point, and each grid point has speed information, so that an input feature matrix X can be obtained2,X2The shape is (m)2×m2,1);
The adjacency matrix A can be obtained by establishing a spatial structure relationship2By means of a adjacency matrixAnddegree matrix ofMixing X2、And substituting the speed information of each grid point into the formula (2), aggregating the speed information of each grid point, predicting to obtain a finite difference coefficient, and calculating a spatial difference and a time derivative.
Further, 3) constructing a graph structure relation for the high-dimensional fluid dynamic equation, wherein the value of each node is represented by the values of a plurality of nodes around the node at the previous time layer, and establishing a spatial structure relation between each node and the nodes around the node, so that all the nodes are spatially associated with fixed nodes.
Further, when training the graph convolutional neural network model, the loss function is:
the loss function uses the mean square error MSE and is the difference L between the estimated time domain difference and the true value1And the difference L between the estimated future state and the actual state2One of these two loss functions is selected to spread the training,
estimated difference L between time-domain difference and true value1Comprises the following steps:
the difference L between the estimated future state and the true state2Comprises the following steps:
wherein m represents m nodes per time layer;representing the time-domain difference at a point estimated by the neural network model at a certain time layer, andthen the true value of the time domain difference at that point is represented; v. ofiRepresenting neural networks at a certain time levelThe magnitude of the velocity at a certain point estimated by the network model, vi' then represents the true value of the velocity at that point;
if finally obtained isUsing a loss function L1To reduce errors; if the velocity value v is to be obtained finally, the loss function L is selected at this time2To optimize model expansion training.
The invention also provides application of the flow velocity prediction method of the self-adaptive high-dimensional fluid dynamic equation based on graph convolution, and the flow velocity prediction method of the self-adaptive high-dimensional fluid dynamic equation based on graph convolution is used for simulating the state of the flow field in the high-pressure pipeline of the nuclear power plant reactor.
The invention also provides application of the flow velocity prediction method of the self-adaptive high-dimensional hydrodynamic equation based on graph convolution.
Compared with the prior art, the invention has the advantages that:
1. when the neural network prediction coefficient is used, the velocity information of the nodes is aggregated by using the information of the edges through graph convolution so as to generate a new node representation, the velocity characteristics (node information) on the grid points and the relation (structure information) between the grid points are fully used, the graph structure relation between the grid points is established, and the graph data takes the node information and the structure information into consideration and fully performs structural representation.
2. And establishing a physical process real relation by fully utilizing the speed characteristics on the grid points and the relation between the grid points. When the graph convolution is used for constructing the graph structure relationship, the incidence relationship of the grid point speed information is fully utilized, the robustness of the speed information is improved, and the rationality of a space mapping mechanism is improved.
3. The method is not only used for a low-dimensional fluid dynamic equation, but also can be used for solving the problem of complex space structure of a high-dimensional fluid dynamic equation, and can be used for predicting, simulating and analyzing a fluid dynamic field, and the method has high calculation speed and high precision.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.
FIG. 1 is a graph in which a velocity value is represented by a velocity value of an upper layer in a one-dimensional case in embodiment 1;
FIG. 2 is a spatial structure diagram obtained in the one-dimensional case of example 1;
FIG. 3 is a graph in which the velocity values are represented by the velocity values of the upper layer in the two-dimensional case in embodiment 1;
FIG. 4 is a graph in which the value of a certain point in the three-dimensional case is represented by the value of the node of the upper time layer in example 1;
FIG. 5 is a schematic flow chart in the one-dimensional case of example 1;
FIG. 6 is a schematic flow chart of the two-dimensional case in example 1.
Detailed Description
The invention is further described with reference to the following figures and specific embodiments.
Example 1
The flow velocity prediction method of the adaptive high-dimensional hydrodynamic equation based on graph convolution is combined with the flow velocity prediction method of the adaptive high-dimensional hydrodynamic equation of the graph convolution and comprises the following steps of:
s1, node x of fluidnThe coarse-grained speed function value is input into a high-dimensional fluid dynamic equation space structure model of the graph convolution neural network, and the model constructs an adaptive space structure relation for each grid point in the grid by establishing a discretization grid and utilizing the graph convolution neural network.
S2, predicting and outputting the space structure coefficient of each grid point by a finite difference-based differential equation numerical method through space structure model calculation
S3, any grid point x obtained by predicting step S2nCoefficient of spatial structure ofAnd xnX of surrounding pointsn-k,…,xn+k(i.e. x)nPoint and xnK points on the left and k points on the right, k being determined by a space structure model) are combined to calculate xnSpatial derivatives of (i.e. x)nProcessing the solution of the velocity value after the spatial evolution; the calculation formula is as follows:
wherein the solution of the formula is a formula in the prior art, and is not described in detail,is the predicted spatial structure coefficient, l is the derivative order, xnRepresented is the coordinate point, v (x)n) Representative is the velocity value at that point, v (x)n-m) Representing the velocity value, x, between two coordinates, point n and point mkRepresenting the accuracy of the difference.
S4, a physical solving process based on the high-dimensional fluid dynamic equation, namely calculating the flux and then calculating x by using the fluxnTime derivative of (i.e. x)nAnd (4) solving the velocity value after the velocity value evolves along with time to finally obtain the final numerical solution of the high-dimensional fluid dynamic equation. Wherein x isnThe flux calculation formula of (b) is:
the time derivative calculation formula is:
the above formula integrates the time derivative to obtain the predicted velocity value, Δ x represents the displacement between the two grids n and m,andare flux values that all refer to half grid points.
In step S1, when constructing the adaptive spatial structure relationship, obtaining the characteristic information of each node, obtaining the characteristic information of the node from all the neighboring nodes of each node, constructing the graph structure relationship, and respectively introducing the specific construction process to the one-dimensional, two-dimensional and high-dimensional fluid dynamic equations as follows;
1) construction of graph structure relationships for one-dimensional fluid dynamic equations
The embodiment adopts a one-dimensional space equation of Burgers, as follows:
10000 time layers are known in the training data set, each layer has 512 grid points, each grid point has its speed information, and 512 grid points are down sampled to obtain 32 grid points. When the model is trained, the velocity values of the 32 grid points of the first layer are used as input, loss is obtained by calculating MAE through the true solution of the velocity values of the next time layer obtained through training and the velocity values of the next time layer, and when the layer is trained, the next layer is trained.
Firstly, a graph structure of a discretization grid is constructed, spatial grid points are used as vertexes of the graph, and relation weights between different grid points are designed to be used as edges of the graphAnd establishing the spatial structure relationship of all grid points. The method comprises the following specific steps: when establishing spatial structure relationship, all m of each layer1The grid points establish a spatial structure relation, the speed value of each grid point is represented by the speed values of 2k grid points around the upper layer and the speed value at the grid point, each grid point has speed information, and therefore an input feature matrix X is obtained according to the speed of each grid point, and the shape of the X is (m)11); since each grid point is represented by 2k points around it, i.e. k points to the left and k points to the right of the point. For the first three points, i.e. P1,P2,P3There are less than k points to the left, so these three points are represented by k points to the right and all points to the left of this point, establishing a spatial structure relationship. Here, k is the optimal number of surrounding points obtained through experiments, that is, the error between the central point speed value obtained through prediction of k surrounding points and the true solution is the minimum.
In this embodiment, when the spatial structure relationship is established, one spatial structure relationship is established for 32 grid points of each layer, and the velocity value of each grid point is represented by the velocity values of 6 grid points around the previous layer and the velocity value at that point, as shown in fig. 1. Each grid point has speed information, so that an input feature matrix X is obtained according to the speed of each grid point, and the shape of the X is (32, 1); the adjacent matrix A can be obtained by establishing a spatial structure relationship, and the adjacent matrix A can be obtained by the adjacent matrixAnddegree matrix ofIs given by the formulaWherein, i is a unit matrix formed by connecting X,And substituting the speed information of each grid point into the formula (2), aggregating the speed information of each grid point, predicting to obtain a finite difference coefficient, and calculating a spatial difference and a time derivative. The following illustrates the specific calculation process:
suppose that the n-th time layer grid point under coarse granularity is 32, and the 32 grid points are respectively P1,P2,P3,...,P31,P32Indicates that the velocity at each grid point is v1,v2,v3,...,v31,v32From which a feature matrix H can be derived(1)(32 × 1) is:
since each grid point is represented by 6 surrounding points, i.e. 3 points to the left and 3 points to the right of the point, for the first three points, i.e. P1,P2,P3There are less than 3 points to the left, so these three points are represented by the 3 points to the right and all points to the left of that point. The finally established spatial structure relationship is shown in fig. 2.
The adjacency matrix a (32 × 32) obtained from the spatial structural relationship is:
can be calculated by the adjacency matrix AAnd degree matrixWherein, i is an identity matrix;is thatDegree matrix (degree matrix) of (1), the formula isObtainedAnd degree matrixRespectively as follows:
and the speed information of the nodes is aggregated by substituting the speed information into the following formula (2), the GCN is also a neural network layer, and the propagation mode between the layers is as follows:
wherein, H is a feature matrix of each layer, obtained from the speed of all grid points of each layer, and is X for the input layer; σ is a nonlinear activation function, which both makes full use of the velocity information of the grid points and their relationship, and enhances the correlation between the grid points.
Second, the machine-learned prediction coefficients can be forced to satisfy the m-th order polynomial constraint, thereby attenuating the approximation error to O (Δ x)m) If the learned discretization is adapted to a smooth solution on the grid scale, the polynomial precision constraint will ensure that the classical finite difference method can be recovered.
2) And (3) constructing a graph structure relation for a two-dimensional fluid dynamic equation:
the two-dimensional spatial equation for Burgers is as follows:
10000 temporal layers are known in the training dataset because there are two dimensions x and y, each layer has 512 × 512 grid points, each grid point has velocity information, and the 512 × 512 grid points are down sampled to obtain 32 × 32 grid points.
A spatial structure relationship is established for 32 × 32 grid points of each layer, and the velocity value of each grid point is represented by the velocity values of 4 grid points around the point and the velocity value of the point in the previous layer, as shown in fig. 3. Each grid point has speed information, so that an input feature matrix X can be obtained2,X2The shape is (32 × 32, 1).
The adjacency matrix A can be obtained by establishing a spatial structure relationship2By means of a adjacency matrixAnddegree matrix ofMixing X2、And substituting the speed information of each grid point into the formula (2), aggregating the speed information of each grid point, predicting to obtain a finite difference coefficient, and calculating a spatial difference and a time derivative.
The specific calculation process is as follows:
assuming that the number of grid points in the nth time layer is 32 × 32 in the coarse granularity in the two-dimensional case, the 32 × 32 grid points are respectively represented by P1,1,P1,2,...,P1,31,P1,32,P2,1,P2,2,...,P2,31,P2,32,...,P32,1,P32,2,...,P32,31,P32,32Indicates that the velocity at each grid point is v1,1,v1,2,...,v1,31,v1,32,v2,1,v2,2,...,v2,31,v2,32,...,v32,1,v32,2,...,v32,31,v32,32From which a feature matrix H can be derived(1)(1024 × 1) is:
the adjacent matrix A can be obtained from the space structure relationship2(1024 × 1024) is:
by means of an adjacency matrix A2Can be calculated to obtain(1024 × 1024) anddegree matrix ofIs given by the formulaObtainedAnd degree matrixRespectively as follows:
and substituting the speed information into the formula (2) to aggregate the speed information of the nodes, predicting to obtain a finite difference coefficient, and calculating a spatial difference and a time derivative.
3) High-dimensional fluid dynamic equation construction diagram structural relation
The one-dimensional and two-dimensional Burgers equations can better show the structural relationship between each time layer and all nodes of one time layer on the time layer by constructing graph convolution space structural relationship, and the method is also suitable for other partial differential equations except the Burgers equation and is also suitable for higher-dimensional partial differential equations except the one-dimensional and two-dimensional equations.
To construct graph structure relationships for high dimensional hydrodynamic equations, for higher dimensional equations, the nodes at each time level are increased many times compared to one and two dimensions due to the increase in dimensions, and the values at each node become more complex when represented by the values of several nodes around the point at the previous time level. In the embodiment, a spatial structure relationship between each node and nodes around the node is established, so that all nodes are spatially associated with fixed nodes, and the value of each node is represented by the values of several nodes around the node at the previous time layer. Fig. 4 is an example of a value represented by a value of several nodes around a point at an upper time layer on a three-dimensional partial differential equation.
In addition, when training the graph convolutional neural network model, the loss function is:
the loss function uses the mean square error MSE and is the difference L between the estimated time domain difference and the true value1(Time derivative loss) and the difference L between the estimated future state and the true state2(Integrated solution loss) selects one of the two loss functions to spread the training,
estimated difference L between time-domain difference and true value1Comprises the following steps:
the difference L between the estimated future state and the true state2Comprises the following steps:
wherein m represents m nodes per time layer;representing the time-domain difference at a point estimated by the neural network model at a certain time layer, andthen the true value of the time domain difference at that point is represented; v. ofiRepresenting the velocity values at a certain point estimated by the neural network model at a certain time layer, and vi' then represents the true value of the velocity at that point.
If finally obtained isUsing a loss function L1To reduce errors; if the velocity value v is to be obtained finally, the loss function L is selected at this time2To optimize model expansion training.
Example 2
This example describes a specific method/application of the present invention for simulating the flow field condition in a high pressure pipeline of a nuclear power plant reactor and analyzing the pressure loss in the pipeline. In the operation process of a nuclear power plant, reactor coolant with high temperature, high pressure and high flow rate in a reactor coolant pipeline (hereinafter referred to as a main pipeline) connecting a reactor pressure vessel and a steam generator transfers heat generated by nuclear fuel in the reactor pressure vessel to the steam generator. At the moment, the flow field state of the reactor needs to be simulated by adopting a plurality of methods, and the pressure loss of the main pipeline is analyzed, so that the proper instruments such as the temperature, the pressure, the flow and the like of the main pipeline are selected, and the reference can be provided for the performance parameters of other related equipment such as a reactor coolant pump and the like.
In the prior art, a finite element method of computational fluid mechanics is generally adopted, a flow variable to be solved is approximated by a simple equation, then the approximate relation is substituted into a continuous control equation to form a discrete equation set, and finally an algebraic equation set is solved. However, the method has the problems that the rough grid division can not be accurately solved, when the grid division is fine, the calculation speed becomes slow, at the moment, the flow velocity prediction method of the self-adaptive high-dimensional fluid dynamic equation based on graph convolution of the invention is utilized, the graph convolution neural network is utilized to establish the incidence relation between the nodes, and the convolution neural network is utilized to replace the calculation of the finite difference coefficient, so that the calculation speed is accelerated, the rationality of space mapping is increased, the steady-state analysis is better carried out on the reactor coolant pipeline between the pressure vessel and the steam generator of the nuclear power plant, so that the fluid state of the reactor coolant pipeline under the normal working condition of the nuclear power plant is simulated (it needs to be noted that the steps of obtaining the flow field state by the method except the prediction speed, and other steps such as a model modeling method and the like can be realized by the prior art), the obtained simulation model can well restore the flow condition of the coolant of the main pipeline, can obtain flow field data richer in theoretical calculation, provides reference for subsequent manufacturing and material selection of main pipeline equipment and model selection design of instruments and control systems, and provides data support for the framework of the control system.
Example 3
This example describes another specific use/application of the present invention for seismic wave numerical simulation, where the velocity value v (x) obtained by applying the method of the present invention is usedn) Is the wave velocity. The seismic wave numerical simulation is an important base stone for seismic exploration, a mathematical theoretical model is formed by effectively simplifying actual problems, the propagation rule of seismic waves in corresponding media is simulated and researched, and seismic records at earth surface or well medium wave detection points are calculated. Therefore, the seismic wave numerical simulation is not only an important tool for understanding the propagation law of the seismic wave and explaining the seismic data, but also a basis of seismic imaging. The seismic wave numerical simulation can simulate acoustic waves, elastic waves, SH waves and the like, and for example, the formula (9) is a two-dimensional isotropic uniform medium acoustic wave equation.
Wherein u isPThe displacement is represented by a displacement of the displacement,the longitudinal wave velocity of the medium is shown, rho is the density of the elastomer, and lambda and mu are Lame coefficients.
Numerical simulation finite difference method, pseudo-spectrum method, finite element method and the like of seismic wave fields. The finite difference method solves the wave equation by using a uniform or non-uniform format, is easy to improve in precision and realize, but can multiply increase the seismic wave simulation time when the model is large in scale and fine in model subdivision. At the moment, by utilizing the scheme provided by the patent, namely the flow velocity prediction method of the graph convolution-based self-adaptive high-dimensional fluid dynamic equation, the graph convolution neural network is used for constructing the space structure relationship and predicting the coefficient, and the other steps still adopt the finite difference method and the prior art, so that not only is the grid division not needed to be fine, but also the seismic waves can be accurately simulated, the simulation time is greatly reduced, and the seismic wave simulation efficiency is improved.
It is understood that the above description is not intended to limit the present invention, and the present invention is not limited to the above examples, and those skilled in the art should understand that they can make various changes, modifications, additions and substitutions within the spirit and scope of the present invention.
Claims (8)
1. The flow velocity prediction method of the self-adaptive high-dimensional hydrodynamic equation based on graph convolution is characterized by comprising the following steps of:
s1, node x of fluidnInputting the coarse-grained speed function value into a high-dimensional fluid dynamic equation space structure model of a graph convolution neural network, and constructing an adaptive space structure relation for each grid point in the grid by establishing a discretization grid by using the graph convolution neural network;
s2, predicting and outputting the space structure coefficient of each grid point by a finite difference-based differential equation numerical method through space structure model calculation
S3, any grid point x obtained by predicting step S2nCoefficient of spatial structure ofAnd xnX of surrounding pointsn-k,…,xn+kThe value of the coarse particle size is combined, i.e. xnPoint and xnK points on the left and k points on the right are adjacent, and x is calculatednSpatial derivatives of (i.e. x)nProcessing the solution of the velocity value after the spatial evolution;
s4, a physical solving process based on the high-dimensional fluid dynamic equation, namely calculating the flux and then calculating x by using the fluxnTime derivative of (i.e. x)nAnd (4) solving the velocity value after the velocity value evolves along with time to finally obtain the final numerical solution of the high-dimensional fluid dynamic equation.
2. The method for predicting the flow velocity of the adaptive high-dimensional fluid dynamic equation based on graph convolution according to claim 1, wherein when the adaptive spatial structure relationship is constructed in step S1, the feature information of each node is obtained, and the feature information of each node is obtained from all neighboring nodes of the node, so as to construct the graph structure relationship, including:
1) construction of graph structure relationships for one-dimensional fluid dynamic equations
Take Burgers' equation as an example. The Burgers equation is a nonlinear partial differential equation that is applied to the field of fluid mechanics to simulate shock wave propagation and reflection. The one-dimensional spatial equation for Burgers is as follows:
Known training data set has T1A time layer, each layer having M1A number of grid points, each having speed information thereon, M1Down sampling at each grid point to obtain m1A plurality of grid points; when training the model, firstly, the m of the first layer is1The speed values of the grid points are used as input, loss is obtained by calculating MAE through the true solution of the speed value of the next time layer obtained through training and the speed value of the next time layer, and when the layer is trained, the next layer is trained;
by constructing a graph structure of a discretization grid, taking spatial grid points as vertexes of the graph, designing relationship weights between every two different grid points as edges of the graph, and establishing spatial structure relationships of all the grid points; the method comprises the following specific steps: when establishing spatial structure relationship, all m of each layer1Establishing a spatial structure relationship at each grid point, wherein the velocity value of each grid point is represented by the velocity values of 2k grid points around the upper layer and the velocity value at the grid point, each grid point has velocity information, and thus the input feature matrix X is obtained from the velocity of each grid point, and the shape of the X is (m)11); since each grid point is represented by 2k points around, i.e. k points to the left and k points to the right of the point; for the first three points, i.e. P1,P2,P3The left side is less than k points, so the three points are represented by the right k points and all the points on the left side of the points, and a space structure relationship is established;
the adjacent matrix A can be obtained by establishing a spatial structure relationship, and the adjacent matrix A can be obtained by the adjacent matrixAnddegree matrix ofIs given by the formulaWherein,i is an identity matrix formed by connecting X,Substituting the speed information into a formula (2), aggregating the speed information of each grid point, predicting to obtain a finite difference coefficient, and calculating a spatial difference and a time derivative;
in addition, the propagation mode between layers of the graph convolution neural network is as follows:
wherein, H is a feature matrix of each layer, obtained from the speed of all grid points of each layer, and is X for the input layer; σ is a nonlinear activation function, which both makes full use of the velocity information of the grid points and their relationship, and enhances the correlation between the grid points.
3. According toThe method of flow velocity prediction for adaptive high-dimensional fluid dynamic equations based on graph convolution of claim 2, wherein the spatial structure relationship is established for m of each layer1The grid points establish a spatial structure relationship, the velocity value of each grid point is represented by the velocity values of the 6 grid points around the previous layer and the velocity value at the point, and since each grid point is represented by 6 points around the grid point, namely 3 points at the left and 3 points at the right of the grid point, the spatial structure relationship is established for the first three points, namely P1,P2,P3The left side is less than 3 points, so the three points are represented by the right 3 points and all the points on the left side of the points, so that the spatial structure relationship is established.
4. The method for flow rate prediction for an adaptive high-dimensional fluid dynamic equation based on graph convolution of claim 2), wherein 2) a graph structure relationship is constructed for a two-dimensional fluid dynamic equation:
the two-dimensional spatial equation for Burgers is as follows:
known training data set has T2A time layer, each layer having M because of two dimensions of x and y1×M1A plurality of grid points, each having speed information, and M1×M1Down sampling at each grid point to obtain m2×m2A plurality of grid points;
when building spatial structure relationship, m for each layer2×m2Establishing a spatial structure relationship at each grid point, wherein the speed value of each grid point is represented by the speed values of 4 grid points around the point at the upper layer and the speed value of the point, and each grid point has speed information, so that an input feature matrix X can be obtained2,X2The shape is (m)2×m2,1);
The adjacency matrix A can be obtained by establishing a spatial structure relationship2By means of a adjacency matrixAnddegree matrix ofMixing X2、And substituting the speed information of each grid point into the formula (2), aggregating the speed information of each grid point, predicting to obtain a finite difference coefficient, and calculating a spatial difference and a time derivative.
5. The method for predicting flow rate of an adaptive high-dimensional fluid dynamic equation based on graph convolution of claim 4, wherein 3) a graph structure relationship is constructed for the high-dimensional fluid dynamic equation, the value of each node is represented by the values of several nodes around the node at the previous time layer, and a spatial structure relationship between each node and the nodes around the node is established, so that all the nodes are spatially associated with a fixed node.
6. The method for flow velocity prediction of adaptive high-dimensional fluid dynamic equations based on graph convolution according to any one of claims 1 to 5, wherein when training the graph convolution neural network model, the loss function is:
the loss function uses the mean square error MSE and is the difference L between the estimated time domain difference and the true value1And the difference L between the estimated future state and the actual state2One of these two loss functions is selected to spread the training,
estimated difference L between time-domain difference and true value1Comprises the following steps:
the difference L between the estimated future state and the true state2Comprises the following steps:
wherein m represents m nodes per time layer;representing the time-domain difference at a point estimated by the neural network model at a certain time layer, andthen the true value of the time domain difference at that point is represented; v. ofiRepresenting the velocity values at a certain point estimated by the neural network model at a certain time layer, and vi' then represents the true value of the velocity at that point;
7. The application of the flow velocity prediction method based on the graph convolution adaptive high-dimensional fluid dynamic equation in claim 6, characterized in that the flow velocity prediction method based on the graph convolution adaptive high-dimensional fluid dynamic equation is used for simulating the flow field state in the high-pressure pipeline of the nuclear power plant reactor.
8. The use of the method for flow velocity prediction of the graph convolution based adaptive high-dimensional hydrodynamic equation according to claim 6, wherein the method for flow velocity prediction of the graph convolution based adaptive high-dimensional hydrodynamic equation is used to construct a spatial structure relationship and predict, obtain seismic wave velocities, and model seismic wavefields.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210300040.2A CN114692523A (en) | 2022-03-25 | 2022-03-25 | Flow velocity prediction method of self-adaptive high-dimensional hydrodynamic equation based on graph convolution |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210300040.2A CN114692523A (en) | 2022-03-25 | 2022-03-25 | Flow velocity prediction method of self-adaptive high-dimensional hydrodynamic equation based on graph convolution |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114692523A true CN114692523A (en) | 2022-07-01 |
Family
ID=82139807
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210300040.2A Pending CN114692523A (en) | 2022-03-25 | 2022-03-25 | Flow velocity prediction method of self-adaptive high-dimensional hydrodynamic equation based on graph convolution |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114692523A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115983148A (en) * | 2022-12-13 | 2023-04-18 | 北京景行锐创软件有限公司 | CFD simulation cloud picture prediction method, system, electronic device and medium |
WO2024040546A1 (en) * | 2022-08-26 | 2024-02-29 | Intel Corporation | Point grid network with learnable semantic grid transformation |
-
2022
- 2022-03-25 CN CN202210300040.2A patent/CN114692523A/en active Pending
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2024040546A1 (en) * | 2022-08-26 | 2024-02-29 | Intel Corporation | Point grid network with learnable semantic grid transformation |
CN115983148A (en) * | 2022-12-13 | 2023-04-18 | 北京景行锐创软件有限公司 | CFD simulation cloud picture prediction method, system, electronic device and medium |
CN115983148B (en) * | 2022-12-13 | 2024-04-12 | 北京景行锐创软件有限公司 | CFD simulation cloud image prediction method, system, electronic equipment and medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Dey et al. | Metamodel based high-fidelity stochastic analysis of composite laminates: A concise review with critical comparative assessment | |
Pfaff et al. | Learning mesh-based simulation with graph networks | |
CN114692523A (en) | Flow velocity prediction method of self-adaptive high-dimensional hydrodynamic equation based on graph convolution | |
CN114724012B (en) | Tropical unstable wave early warning method and device based on space-time cross-scale attention fusion | |
CN108763683A (en) | New WENO format building methods under a kind of trigonometric function frame | |
Das et al. | State-of-the-art review of design of experiments for physics-informed deep learning | |
CN114611350A (en) | FCN-based phase change heat storage device flow channel topology optimization design method | |
CN110705183A (en) | Multilayer grid LBM evolution method with buffer area | |
Drakoulas et al. | FastSVD-ML–ROM: A reduced-order modeling framework based on machine learning for real-time applications | |
Zhang et al. | A hybrid sequential sampling strategy for sparse polynomial chaos expansion based on compressive sampling and Bayesian experimental design | |
CN103530451B (en) | Many grids Chebyshev parallel spectral element method of complex dielectrics elastic wave modeling | |
CN117854643B (en) | MEMS membrane simulation method and system based on graphic neural network | |
Xu et al. | A novel ensemble model using artificial neural network for predicting wave-induced forces on coastal bridge decks | |
Hart et al. | An advanced cost estimation methodology for engineering systems | |
Zou et al. | Deep neural Helmholtz operators for 3D elastic wave propagation and inversion | |
CN115879335A (en) | Fluid multi-physical-field parameter prediction method based on graph-generated neural network | |
CN117556713A (en) | Uncertainty quantization method for CFD multi-credibility high-dimensional correlation flow field | |
CN117057265A (en) | Flow field prediction method based on migration learning and discrete physical information neural network | |
CN115758853A (en) | Reservoir facies modeling method based on sequential neural network | |
Shahane et al. | Convolutional neural network for flow over single and tandem elliptic cylinders of arbitrary aspect ratio and angle of attack | |
Silva et al. | GAN for time series prediction, data assimilation and uncertainty quantification | |
Boujoudar et al. | Implicit EXP-RBF techniques for modeling unsaturated flow through soils with water uptake by plant roots | |
CN103971411A (en) | Space curved surface modeling method by utilizing space curved surface sampling points of three-dimensional objects | |
Ye et al. | PEPNet: A barotropic primitive equations-based network for wind speed prediction | |
Fujita et al. | Scalable finite-element viscoelastic crustal deformation analysis accelerated with data-driven method |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |