Fast prediction method for complex pipe network system with multi-physical fields based on element agent model

Abstract

This invention discloses a rapid multiphysics prediction method for complex pipe network systems based on a component proxy model, comprising the following steps: constructing a one-dimensional pipe network system model based on a three-dimensional model of the actual ventilation pipe network system; constructing a component-level proxy model based on the constructed one-dimensional pipe network system model using a neural operator method; training and optimizing the model using the DeepONet method based on the neural operator method; and substituting the constructed component proxy model into the one-dimensional pipe network system model for calculation, thereby achieving multiphysics prediction from the component level to the system level, optimizing system performance, and selecting the optimal option. This invention, employing the aforementioned rapid multiphysics prediction method for complex pipe network systems based on a component proxy model, solves the problems of long computation time and large resource consumption in three-dimensional computation of complex systems, significantly reducing the time and resources required for traditional CFD simulation, and achieving more efficient and accurate system analysis and optimization.

Description

Technical Field

[0001] This invention relates to the fields of computational fluid dynamics and pipeline systems, and in particular to a method for rapid multiphysics prediction of complex pipeline systems based on component proxy models. Background Technology

[0002] In modern buildings and industrial facilities, the design and operational efficiency of pipeline systems directly impact indoor air quality, energy consumption, and environmental comfort. With the continuous increase in building scale and the diversification of functions, traditional pipeline systems face growing challenges, particularly in the design and optimization process. A key issue is how to efficiently and accurately calculate and simulate the coupling effects of multiple physical fields, such as airflow, pressure distribution, and temperature changes.

[0003] Currently, the calculation of pipeline systems mainly relies on three-dimensional numerical simulation, especially the use of computational fluid dynamics (CFD) methods to simulate airflow and heat transfer. CFD methods can provide high-precision fluid dynamics calculation results and are widely used in areas such as fan performance, pipeline design, and heat conduction analysis. However, the application of CFD methods typically requires substantial computational resources and lengthy calculation processes, especially when dealing with large-scale or complex pipeline systems, where computation time and hardware resource consumption increase exponentially. Therefore, traditional three-dimensional simulation methods are not efficient enough for large-scale systems, especially when rapid, iterative optimization design is required, making it difficult to meet real-time requirements.

[0004] Furthermore, in the actual design and operation of pipeline systems, calculations often cannot be completed at the system level in one go, especially for complex building and industrial facility pipeline systems, which often require calculations by region or level. While such zonal calculations can improve computational efficiency to some extent, they are also prone to introducing computational errors, particularly in the coupling between multiple regions and the handling of boundary conditions, which often leads to inaccurate calculation results and underestimation of system performance.

[0005] In summary, while existing 3D CFD calculation methods are accurate, their computational complexity and resource consumption are too high, making it impossible to quickly meet design and optimization needs. Therefore, how to reduce computational complexity, improve computational efficiency, and maintain accuracy has become a technical challenge to be solved. Summary of the Invention

[0006] The purpose of this invention is to provide a rapid multiphysics prediction method for complex pipeline systems based on a component proxy model. By modeling each component in the pipeline system as an assembly and connecting these components through a unified one-dimensional system, the calculation process of traditional three-dimensional models is simplified, while retaining key physical data and effectively interacting with the three-dimensional system. This method not only reduces computational load but also improves the system's computational efficiency and accuracy, making it particularly suitable for complex pipeline systems requiring rapid calculation and optimization.

[0007] To achieve the above objectives, this invention provides a fast multiphysics prediction method for complex pipeline systems based on a component proxy model, comprising the following steps:

[0008] S1. Construct a one-dimensional duct system model based on the three-dimensional model of the actual ventilation duct system;

[0009] S2. Based on the components in the one-dimensional pipeline system model constructed in S1, construct a component-level proxy model to realize one-dimensional and three-dimensional interaction at the component level;

[0010] S3. Replace all components in the one-dimensional pipe network system with the component-level proxy model constructed in S2 to form a system-level multiphysics coupling.

[0011] Preferably, in S1, a one-dimensional pipe network system model is constructed, including the flow and heat transfer characteristic control equations and the boundary conditions of the multi-physics field characteristics to describe the flow and heat transfer process of the fluid in the pipe.

[0012] The governing equations for flow heat transfer characteristics include the Navier-Stokes equations, the continuity equation, and the energy conservation equation. The Navier-Stokes equations are as follows:

[0013]

[0014] The continuity equation is:

[0015]

[0016] Where ρ is the fluid density; t is time; This is the Nabla operator, used for gradient and divergence calculations; It is the velocity vector; p is the pressure; μ is the kinematic viscosity coefficient; For physical strength;

[0017] The energy conservation equation is:

[0018]

[0019] Where e is the internal energy per unit mass; T is the temperature; k is the thermal conductivity; Φ is the viscous power dissipation; and Q is the volumetric heat source.

[0020] The boundary conditions for multiphysics characteristics include the initial conditions of the components and the flow and heat transfer boundary conditions.

[0021] Preferably, in S2, constructing the component-level proxy model includes the following steps:

[0022] Sa1. Determine the input parameters of the component-level proxy model and the output parameters that need to be returned to the one-dimensional system. The input parameters include geometry, flow rate, and pressure; the output parameters include pressure and flow rate.

[0023] Sa2. Conduct sampling design and select and build the model. Based on the distribution of sample points, select Latin hypercube sampling, random sampling, or sequential sampling methods for sampling design.

[0024] Sa3, training data through multiphysics characteristics;

[0025] Sa4. Model validation and evaluation, with evaluation metrics including mean squared error (MSE) and coefficient of determination (R²). 2 The loss function is optimized using the Adam optimizer based on the mean absolute error (MAE) and the results of the optimization are then applied.

[0026] Preferably, in S2, the methods for constructing the component-level proxy model include, but are not limited to, autoencoders, neural operators, POD reduction models, and DMD reduction models.

[0027] Preferably, in S2, a component-level surrogate model is constructed using a neural operator method. The DeepONet method is used to learn the mapping relationship from input to output based on a neural network. The specific process is as follows:

[0028] Sb1. Prepare a dataset containing the input function u(x) and the output function v(x), where each sample contains the values ​​u(x) of the input function at multiple spatial points. i ) and output function v(x i );

[0029] Sb2. Construct a Branch Net, using the function u(x) as the input to the Branch Net. u(x) is a physical quantity at different positions (x1, x2, x3...x...). m The values ​​of ) are [u(x1), u(x2), ..., u(x)]. m The function u(x) is mapped to a high-dimensional latent space, and the output is a vector. This represents the characteristics of each point in the latent space;

[0030] Sb3. Construct a Trunk Net, taking spatial coordinates x∈D as input, mapping the spatial coordinates x to another latent space, and outputting a vector. This indicates the spatial characteristics of the location;

[0031] Sb4, the output function G(u)(y) is passed through b k and t k The inner product or dot product is generated as follows:

[0032]

[0033] Prediction result u pred,i With label data u gt,i The mean squared error (MSE) is used as the loss function to measure the difference between the predicted result and the true value.

[0034]

[0035] Where N represents the number of samples;

[0036] The Adam optimization algorithm is used to adjust the parameters in the network, and the model is optimized by minimizing the loss function to evaluate the model's accuracy and stability.

[0037] Therefore, the present invention employs the above-mentioned fast multiphysics prediction method for complex pipeline systems based on the component proxy model, and the beneficial effects are as follows:

[0038] (1) By constructing a one-dimensional pipeline network system and component proxy model, this invention simplifies the calculation of traditional three-dimensional models, reduces the amount of calculation, significantly shortens the calculation time, and enables rapid multiple iterations of optimization design to meet real-time requirements.

[0039] (2) This invention uses neural operators to construct a component-level proxy model, which can accurately learn the intrinsic laws of multi-physics field functions. Combined with DeepONet training and optimization, it can effectively improve the accuracy and stability of the model and achieve accurate multi-physics field prediction.

[0040] (3) This invention constructs proxy models for different component structures and operating conditions, which are highly adaptable and can be widely applied to HVAC systems, industrial pipeline systems and other fields, helping to optimize the design, analysis and operation of pipeline systems in various fields.

[0041] (4) The present invention constructs a system-level proxy model based on a component-level proxy model to realize the multi-physics coupling of the system. In actual working conditions, the flow field in the component can be viewed under any circumstances, which provides great convenience for engineering research.

[0042] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0043] Figure 1This is an embodiment of the multiphysics fast prediction method for complex pipeline systems based on the component proxy model of the present invention, which describes the pipeline system model construction process, wherein (a) is a schematic diagram of the three-dimensional pipeline system model structure, and (b) is a schematic diagram of the one-dimensional pipeline system constructed based on the three-dimensional pipeline structure, and the component proxy model construction process therein;

[0044] Figure 2 This is a flowchart illustrating the computational fluid dynamics parameter calculation process of an embodiment of the multiphysics-based rapid prediction method for complex pipeline systems based on a component proxy model, according to the present invention.

[0045] Figure 3 This is a schematic diagram of the DeepONet method based on BranchNet and TrunkNet, an embodiment of the fast multiphysics prediction method for complex pipeline systems based on the component proxy model of the present invention. Detailed Implementation

[0046] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0047] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0048] This invention first constructs a proxy model of the components, and then forms a system through the various components to construct a system-level proxy model for calculation. This method can be widely applied in fields such as HVAC systems and industrial pipeline systems to optimize the calculation process in design, analysis and operation.

[0049] As shown in the figure, the method for rapid multiphysics prediction of complex pipeline systems based on the component proxy model includes the following steps:

[0050] S1, such as Figure 1 As shown in (a), based on the three-dimensional model of the actual ventilation duct network system, the geometric parameters of each component in the duct network system (such as the diameter of the fan, the length and diameter of the pipe, etc.) are obtained; each component (fan, pipe, valve, pump, etc.) is modeled according to its geometric features and fluid dynamic system; the components are connected through a one-dimensional calculation system to form a complete and simplified duct network system.

[0051] like Figure 1 As shown in (b), a complete one-dimensional pipeline system model is constructed, in which all the components in the three-dimensional system are constructed in one dimension, and the upstream and downstream are connected according to the actual working conditions to form a complete pipeline system; specifically, it includes the flow and heat transfer characteristic control equations and the boundary conditions of the multi-physics field characteristics, which are used to describe the flow and heat transfer process of fluid in the pipeline.

[0052] The governing equations for flow heat transfer characteristics include the Navier-Stokes equation, the continuity equation, and the energy conservation equation. The Navier-Stokes equation is as follows:

[0053]

[0054] The continuity equation is:

[0055]

[0056] Where ρ is the fluid density; t is time; This is the Nabla operator, used for gradient and divergence calculations; It is the velocity vector; p is the pressure; μ is the kinematic viscosity coefficient; For physical strength.

[0057] The energy conservation equation is:

[0058]

[0059] Where e is the internal energy per unit mass; T is the temperature; k is the thermal conductivity; Φ is the viscous power dissipation; and Q is the volumetric heat source.

[0060] The boundary conditions for multiphysics characteristics include the initial conditions of the components and the flow and heat transfer boundary conditions.

[0061] S2. Based on the one-dimensional pipeline system model constructed in S1, a component-level proxy model is built for the components to realize one-dimensional and three-dimensional interaction at the component level.

[0062] This embodiment mainly describes a method for constructing a high-precision adaptive component-level surrogate model of the multi-physics characteristics of complex components, including but not limited to constructing autoencoders, neural operators, POD reduced-order models, and DMD reduced-order models.

[0063] An autoencoder is an unsupervised learning model that aims to learn a low-dimensional "latent representation" from high-dimensional input data and reconstruct the original input through the representation. It generally consists of an encoder, a latent space / code layer, a decoder, and a loss function.

[0064] POD and DMD are two different algorithms for order reduction models. The core idea of ​​the POD model is to extract the "linear" principal structure using orthogonal bases. The core idea of ​​the DMD model is to fit a linear evolution operator onto a time-series snapshot to extract the space-time modes. All three methods can be used to construct reduced-order models; the choice depends on the specific needs.

[0065] like Figure 2As shown, constructing a component-level proxy model specifically includes the following steps:

[0066] Sa1. Determine the input parameters of the component-level proxy model and the output parameters that need to be returned to the one-dimensional system. The input parameters include geometry, flow rate, and pressure; the output parameters include pressure and flow rate.

[0067] Sa2. Conduct sampling design and select and build the model. Based on the distribution of sample points, select Latin hypercube sampling, random sampling, or sequential sampling methods for sampling design.

[0068] Sa3, training data through multiphysics characteristics;

[0069] Sa4. Model validation and evaluation, with evaluation metrics including mean squared error (MSE) and coefficient of determination (R²). 2 The loss function is optimized using optimizers such as Adam, and then applied after optimization and improvement based on the verification results.

[0070] A component-level surrogate model is constructed using neural operator methods. Through the DeepONet method, the mapping relationship from input to output is learned based on a neural network, particularly when both input and output are functions. The specific process is as follows:

[0071] Sb1, such as Figure 3 As shown in (A) and (B), prepare a dataset with an input function u(x) and an output function v(x). Each sample contains the values ​​of the input function u(x) at multiple spatial points. i ) and output function v(x i ).

[0072] Sb2, such as Figure 3 As shown in (C), a Branch Net is constructed, with the function u(x) as the input to the Branch Net. u(x) represents the physical quantity (temperature, velocity, etc.) at different locations (x1, x2, x3, ... x). m The values ​​of ) are [u(x1), u(x2), ..., u(x)]. m The input function u(x) is mapped to a high-dimensional latent space, and the output is a vector. This represents the characteristics of each point in the latent space.

[0073] Sb3, such as Figure 3 As shown in (D), a Trunk Net is constructed. The input spatial coordinates x∈D are mapped to another latent space, and the output is a vector. This indicates the spatial characteristics of the location.

[0074] Sb4, the output function G(u)(y) is passed through bk and t k Inner product or dot product generation:

[0075]

[0076] Prediction result u pred,i With label data u gt,i The mean squared error (MSE) is used as the loss function to measure the difference between the predicted result and the true value.

[0077]

[0078] Where N represents the number of samples;

[0079] The Adam optimization algorithm is used to adjust the parameters in the network, and the model is optimized by minimizing the loss function to evaluate the model's accuracy and stability.

[0080] S3. Replace all components in the one-dimensional pipeline system constructed in S1 with the component-level proxy model constructed in S2 to form a system-level multiphysics coupling, realize multiphysics prediction from the component level to the system level, and quickly evaluate and optimize the system performance to select the optimal option.

[0081] Therefore, this invention adopts the above-mentioned method for rapid prediction of multiphysics fields in complex pipe network systems based on component proxy models. It establishes proxy models of components based on neural operator proxy models, and calculates multi-dimensional (one-dimensional, three-dimensional) multiphysics fields through the component proxy models in a one-dimensional ventilation system. At the same time, it uses neural operators to establish proxy models for CFD, which can significantly reduce the time and resources required for traditional CFD simulation and achieve rapid prediction.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A fast multiphysics prediction method for complex pipeline systems based on a component proxy model, characterized in that, Includes the following steps: S1. Construct a one-dimensional duct network system model based on the three-dimensional model of the actual ventilation duct network system; S2. Based on the components in the one-dimensional pipeline system model constructed in S1, construct a component-level proxy model to realize one-dimensional and three-dimensional interaction at the component level; S3. Replace all components in the one-dimensional pipe network system with the component-level proxy model constructed in S2 to form a system-level multiphysics coupling. In S1, a one-dimensional pipe network system model is constructed, including the flow and heat transfer characteristic control equations and the boundary conditions of the multi-physics field characteristics to describe the flow and heat transfer process of fluid in the pipe. The governing equations for flow heat transfer characteristics include the Navier-Stokes equations, the continuity equation, and the energy conservation equation. The Navier-Stokes equations are as follows: ； The continuity equation is: ； in, For fluid density; For time; This is the Nabla operator, used for gradient and divergence calculations; It is a velocity vector; Pressure; The coefficient of kinematic viscosity; For physical strength; The energy conservation equation is: ； in, Internal energy per unit mass; For temperature; Thermal conductivity; For viscous dissipation power; As a volumetric heat source; The boundary conditions for multiphysics characteristics include the initial conditions of the components and the flow and heat transfer boundary conditions.

2. The method for rapid multiphysics prediction of complex pipeline systems based on a component proxy model according to claim 1, characterized in that, In S2, constructing a component-level proxy model involves the following steps: Sa1. Determine the input parameters of the component-level proxy model and the output parameters that need to be returned to the one-dimensional system. The input parameters include geometry, flow rate, and pressure; the output parameters include pressure and flow rate. Sa2. Conduct sampling design and select and build the model. Based on the distribution of sample points, select Latin hypercube sampling, random sampling, or sequential sampling methods for sampling design. Sa3, training data through multiphysics characteristics; Sa4. Model validation and evaluation, with evaluation metrics including mean squared error (MSE) and coefficient of determination (R²). 2 The loss function is optimized using the Adam optimizer based on the mean absolute error (MAE) and the results of the optimization are then applied.

3. The method for rapid multiphysics prediction of complex pipeline systems based on a component proxy model according to claim 2, characterized in that, In S2, methods for constructing component-level proxy models include, but are not limited to, autoencoders, neural operators, POD reduction models, and DMD reduction models.

4. The method for rapid multiphysics prediction of complex pipeline systems based on a component proxy model according to claim 3, characterized in that, In S2, a neural operator method is used to construct a component-level surrogate model. The DeepONet method is used to learn the mapping relationship from input to output based on a neural network. The specific process is as follows: Sb1, Preparing the Input Function and output function The dataset contains the values ​​of the input function at multiple spatial points for each sample. and output function ; Sb2, Construct a Branch Net and assign functions As input to the Branch Net, It is a physical quantity, in different locations The value of , input function Mapped to a high-dimensional latent space, the output is a vector. , representing the characteristics of each point in the latent space; Sb3, Construct a Trunk Net and input spatial coordinates. spatial coordinates Mapped to another latent space, the output is a vector. This indicates the spatial characteristics of the location; Sb4, Output Function pass and The inner product or dot product is generated as follows: ； Prediction results With tag data The mean squared error (MSE) is used as the loss function to measure the difference between the predicted result and the true value. ； in, Indicates the number of samples; The Adam optimization algorithm is used to adjust the parameters in the network, and the model is optimized by minimizing the loss function to evaluate the model's accuracy and stability.

Images