Hydrogen shaft furnace gas-solid coupling modeling method based on physical information neural network enhancement

Abstract

The application provides a hydrogen shaft furnace gas-solid coupling modeling method based on physical information neural network enhancement. For the solving part of the NS equation required for gas phase motion modeling, the hybrid function part of the SST k‑ ω turbulence model is enhanced by using Neural ODE based on a large amount of historical data of the hydrogen shaft furnace, so that the solving accuracy of the SST k‑ω turbulence model in different regions of the furnace is improved; for the solid phase motion modeling part, the discrete element method (DEM) is used to calculate the force between solid particles, and Newton's law is used to model the motion equation of the solid particles; for the coupling modeling part, the solid phase motion equation and the gas phase motion equation are coupled based on the drag force model. The application utilizes the complex fitting capability of PINN, so that we can more accurately switch the turbulence model between different regions, and then more accurate results can be obtained for the internal gas-solid coupling motion modeling of the hydrogen shaft furnace.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of gas-solid two-phase motion modeling in hydrogen shaft furnace, and particularly relates to a hydrogen shaft furnace gas-solid coupling modeling method based on physical information neural network enhancement. BACKGROUND

[0002] As an advanced metallurgical equipment, hydrogen shaft furnace realizes the goal of clean production and energy saving and emission reduction by reducing iron ore with hydrogen. However, the gas-solid two-phase flow process in the shaft furnace is very complex, involving turbulent gas motion, interaction between particles, and gas-solid coupling and other physical phenomena. These processes interact with each other, not only have an important influence on the operating efficiency and stability of the shaft furnace, but also directly determine the reduction effect and energy utilization efficiency of the shaft furnace. Therefore, how to accurately model and simulate the gas-solid coupling motion process in the hydrogen shaft furnace has become a key technical problem for improving the performance of the shaft furnace.

[0003] For the modeling of gas phase motion, a turbulent flow model is usually used, and the SST k-ω model is widely used for flow field simulation because it can better handle the behavior of turbulent boundary layers. The SST k-ω model combines the advantages of k-ω model and k-∈ model, and realizes the dynamic switching of the two models through a mixing function F1, thereby improving the adaptability to complex flow fields. However, the model still has certain limitations in dealing with multi-physical field coupling problems. The various parameters in the mixing function F1 depend on empirical formulas, lack the ability to adaptively adjust the dynamic changes of turbulent characteristics, and the switching precision of the model between different regions is limited, making it difficult to meet the accurate description needs of the complex flow in the shaft furnace.

[0004] Secondly, in the modeling of solid particles, the collision, contact and friction between solid particles in the shaft furnace significantly affect the motion trajectory and accumulation morphology of the particles. The discrete element method is currently the mainstream method for studying the motion of solid particles. By calculating the interaction forces between particles, the particle motion can be accurately simulated. However, the complexity of gas-solid coupling motion in the shaft furnace makes it difficult to fully reflect the interaction between particles and flow field by relying solely on the DEM method. Especially under the action of drag force, the gas flow field will significantly affect the motion state of the particles, and the motion of the particles will in turn change the distribution of the gas flow field, thereby leading to the nonlinear enhancement of the gas-solid interaction.

[0005] In addition, in the gas-solid coupling modeling, how to efficiently couple the gas phase turbulent flow model and the solid phase particle model is a key technical problem. The drag force model is usually used to describe the force of the gas phase on the particles, but the drag force coefficient C d , the relative velocity U relThe equal parameters are affected by the complexity of the flow field, and are difficult to accurately estimate by traditional empirical formulas. Meanwhile, the motion state of the particles will exert feedback on the gas phase flow field through the reaction force, and this two-way coupling process needs to be solved iteratively in numerical simulation, which is low in calculation efficiency and is easily affected by boundary conditions and initial value settings, leading to instability of the solution. SUMMARY

[0006] In order to overcome the shortcomings of the existing method, the purpose of the present application is to provide a hydrogen shaft furnace gas-solid coupling modeling method based on physical information neural network enhancement. The present application first uses a parameterization method to parameterize and calculate multiple variables of the mixing function in the SST k-ω turbulence model, and uses Neural ODE for fitting. On the basis of the above parameters, the PINN is used to build the calculation framework of the SST k-ω turbulence model to realize the complete model training process. Through the above accurate training of the deep neural network, the mixing function of the turbulence model is enhanced, which can better replace the traditional empirical method and realize the switching of the more smooth calculation mode. On the basis of the above turbulence model, combined with the discrete element method and the drag force model, the gas-solid coupling motion modeling of the gas and solid particles in the shaft furnace is realized, which can be widely applied to the complex coupling gas-solid coupling motion modeling scene.

[0007] A hydrogen shaft furnace gas-solid coupling modeling method based on physical information neural network enhancement, the steps are:

[0008] Step 1: For the mixing function of the SST k-ω model, use a neural network-based dynamic adjustment mechanism to convert the parameters of the mixing function into a differentiable Neural-ODE network form;

[0009] Step 2: From a large number of collected hydrogen shaft furnace data related to gas flow and furnace operating state, extract and normalize variables, and input these preprocessed data into the Neural ODE network to output the core variables W and θ of the enhanced mixing function;

[0010] Step 3: Establish a deep neural network-based SST k-ω model calculation framework, build a network to predict the hydrogen velocity field and pressure field, and simultaneously use the turbulence kinetic energy and turbulence frequency k and ω as key variables associated with spatial position and time;

[0011] Step 4: Use the output velocity field and pressure field of the network, as well as the corresponding differential relationship, to build a total loss function including initial condition loss, boundary condition loss and control equation residual loss;

[0012] Step 5: Minimize the total loss function by the Adam optimization algorithm to iteratively update the network parameters to reduce the error;

[0013] Step six: after the optimization is completed, the trained PINN network is used to directly propagate the fluid flow results, and the actual collected data is used as a benchmark to output the final velocity and pressure field distribution;

[0014] Step seven: based on the velocity and pressure field distribution output by the PINN network, further combined with solid phase modeling, the discrete element method is used to calculate the force between solid particles, and the solid particle motion equation is established based on Newton's law, the motion equations of gas and solid phases are coupled through the drag force model, and the gas-solid coupling solution is realized.

[0015] In step one, the mixing function of the SST k-ω model is enhanced, first the mixing function is defined and parameterized, and the fluid motion control equation of the over-parameterized SST k-ω model is as follows:

[0016]

[0017] Where ρ and u represent the density and velocity component of the fluid, k and ω represent the turbulent kinetic energy and turbulent frequency, P k represents the generation term of turbulent kinetic energy, which represents the turbulent kinetic energy generated by the shear action of the average flow field, as a correction term to avoid numerical instability; F1 is the mixing function after enhancement, which is used for smooth transition between k-ω and k-∈ model, the function expression is as follows:

[0018]

[0019] In the above formula, α, β, σ ω are empirical constants, which can be fine-tuned according to the situation, y is the normal distance between the calculation area and the wall, v is the kinematic viscosity, σ k and σ ω are turbulent viscosity diffusion coefficients, W and θ are enhancement coefficients to enhance the dynamic fitting ability of the mixing function, and Neural ODE is used for fitting the parameterized mixing function; the equation of F1 mixing function using Neural ODE is as follows:

[0020]

[0021] Where f is a nonlinear function representing the differential term, a linear layer network is used here, and the activation function in the middle uses the Tanh function; the input part of the model here is the turbulent kinetic energy, turbulent frequency, wall distance and the parameters involved in the above mixing equation, the enhancement parameters W, θ are brought into the mixing equation to obtain the parameterized mixing function, and finally the parameterized calculation result of F1 is obtained, which is used for subsequent model calculation; the data that the shaft furnace can record includes hydrogen flow, temperature and pressure, before the Neural ODE model calculation, k, ω are calculated according to the hydrogen flow and temperature,

[0022]

[0023] where I = 0.16Re -1 / 8 , Re is the Reynolds number, C μ is a constant, taken as 0.09, and L is the turbulent length scale, taken as 0.07 of the pipe diameter; in step two, the data of the actual production site are processed to select variables and outliers, to obtain variables related to hydrogen flow, including hydrogen flow, temperature, pressure, and to standardize the selected variables, to generate time series data using a sliding window as input data for the overall Neural ODE, and the model outputs W and θ.

[0024] In step three, an SST k-ω model calculation framework based on a deep neural network is established, and the input and output variables and the overall calculation process of the framework are set according to the actual operation of the shaft furnace; a linear layer network is used to build an SST k-ω-based NS equation calculation framework, the model input uses the spatial motion coordinates of hydrogen and the recorded time steps, the model output of each unit is assumed to represent the velocity field distribution (u, v, w) and the pressure field distribution p at different positions, and the partial derivatives with respect to the spatial coordinates and time are directly solved based on the automatic differentiation technique; after obtaining the output results of the model, the overall model training loss is designed according to step four, the velocity, pressure and corresponding differential components output by the linear layer neural network are used to construct a loss function for PINN network training; the overall loss function includes six items, namely the initial condition loss, the boundary condition loss, the momentum loss term, the mass loss term and the k and ω loss terms; for the initial condition loss, the difference between the velocity distribution of the initial flow field and the velocity distribution of the model output flow field is directly calculated, and the formula is as follows

[0025] L init =||u(x,y,z,t=0)-u i (x,y,z)| 2 (7); For the boundary condition loss, it is composed of two parts, namely the Dirichlet boundary condition loss term and the Neumann boundary condition loss term, wherein the Dirichlet boundary condition loss term is represented as:

[0026] L BC-Dirichlet =||u(x,y,z,t)-u b (x,y,z,t)| 2 (8);

[0027] where u(x,y,z,t) is the velocity field distribution at different positions at time t output by the model, and u b(x,y,z,t) represents the velocity field distribution at different locations at time t under actual operating conditions; the Neumann boundary condition loss term is expressed as...

[0028]

[0029] The purpose of this loss is to ensure that the directional derivative of the velocity field output by the model is zero in the normal direction at the boundary locations, i.e., there is no shear flow. The loss for the governing equations is divided into the momentum equation loss and the mass equation loss, as follows:

[0030]

[0031] To further constrain the model, a constraint term is introduced into the SST k-ω equation, as shown in the following formula:

[0032]

[0033] L k =||R k || 2 (14);

[0034] L ω =||R ω || 2 (15); After synthesis, the training loss of the entire PINN is

[0035] L = L init +L bC-Neumann +L bC-Dirichlet +L momentum +L continuity +L k +K ω (16); After obtaining the loss, as described in step five, the total loss function is minimized using the Adam optimization algorithm, and the network weights are updated using backpropagation to gradually reduce the loss value. During training, the decreasing trend of the total loss and its components is monitored periodically to ensure that the model's predicted solution meets physical constraints and boundary conditions. The velocity distribution of the flow field is directly output using the trained PINN network, and gradient backpropagation is performed based on the actual collected data as labels, using the overall loss mentioned above. Finally, in actual modeling, the velocity distribution of the flow field is directly output using the trained Neural ODE and PINN networks, and only the parameters W and θ output by the Neural ODE are used. These parameters are then substituted into the SST k-ω turbulence model to model the gas motion.

[0036] In step seven, based on the velocity field and pressure field distribution output by PINN, the solid phase motion is modeled combined with DEM, and the gas-solid motion is combined based on the drag force model;Discrete element method independently analyzes the force condition of each particle, and models the particle motion equation according to Newton equation, the force condition and motion equation are as follows:

[0037]

[0038] Where F gp is the pressure gradient force, F drag is the drag force, F c is the collision pressure between particles, which is divided into normal force and tangential force, represent the tangential contact force between particle i and particle j;After considering the drag force in the solid phase motion modeling equation, the drag force term is considered in the gas phase motion, which has an influence on the turbulent kinetic energy and turbulent frequency respectively, and the formula result is as follows

[0039]

[0040] Where is the gradient of gas velocity field, F drag is the drag force, U rel = U g -v p , which represents the relative velocity of gas phase and solid phase, and γ is the model constant related to drag force, U rel = U g -v p , C d : drag force coefficient, related to Reynolds number;By introducing the drag force into the turbulent flow equation, the distribution of k and ω is corrected, and the updated parameters are used to correct the velocity field distribution, and finally the gas-solid coupling motion modeling is realized.

[0041] The present application has the beneficial effects of:

[0042] 1、By introducing the dynamic adjustment mechanism based on neural network, the mixing function in SST k-ω model is parameterized modeling, and the differentiable expression of mixing function is realized by Neural-ODE, which greatly enhances the adaptive adjustment ability of turbulent model. This method can dynamically optimize the distribution of turbulent kinetic energy and turbulent frequency, and shows higher precision in complex flow field.

[0043] 2、Based on the velocity field and pressure field output by PINN, the accurate modeling of the force and motion between solid particles is realized by combining with discrete element method. Through the coupling of gas phase and solid phase motion equation by drag force model, a two-way gas-solid coupling model is constructed, which effectively captures the force of gas phase on solid phase and the feedback effect of solid phase on gas phase.

[0044] 3. By accurately simulating the gas-solid coupling motion process, the method can optimize the flow field distribution in the shaft furnace, improve the particle reduction efficiency, reduce energy waste and carbon emissions, and provide a strong guarantee for the promotion of green metallurgical technology. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 is a model architecture diagram of the hydrogen shaft furnace gas-solid coupling modeling method based on physical information neural network enhancement of the present application.

[0046] Figure 2 is a whole solution flow chart of the hydrogen shaft furnace gas-solid coupling modeling method based on physical information neural network enhancement of the present application. DETAILED DESCRIPTION

[0047] As Figure 1 , Figure 2 illustrated, the implementation steps of the present application are described in detail.

[0048] I. Parameterized SST k-ω turbulence model

[0049] In the current step, the SST k-ω model mixing function needs to be enhanced. First, the mixing function needs to be defined and parameterized. The fluid motion control equation of the parameterized SST k-ω model is as follows:

[0050]

[0051] wherein, as a correction term to avoid numerical instability. F1 is the enhanced mixing function used for smooth transition between k-ω and k-∈ models. The specific function expression is as follows:

[0052]

[0053] In the above formula, α, β, σ ω are empirical constants, which can be fine-tuned according to the situation, y is the normal distance between the calculation region and the wall, v is the kinematic viscosity, σ ω is the turbulent viscosity diffusion sparse, W and θ are enhancement coefficients to enhance the dynamic fitting ability of the mixing function. In this method, Neural ODE is used for fitting the parameterized mixing function. The specific equation of F1 mixing function using Neural ODE is as follows:

[0054]

[0055] where f is a nonlinear function representing the differential term, a linear layer network is used here, and the activation function in the middle uses the Tanh function. The input part of the model here is the turbulent kinetic energy, turbulent frequency, wall distance, and the above-mentioned mixing equation involving parameters, and the enhanced parameters W, θ are brought into the mixing equation to obtain the parameterized mixing function, and finally the parameterized calculation result of F1 is obtained, which is used for subsequent model calculation. However, the data that the actual shaft furnace can record only contains hydrogen flow, temperature and pressure, so before the Neural ODE model calculation, k, ω,

[0056]

[0057] where I = 0.16Re -1 / 8 , Re is the Reynolds number, C μ is a constant, usually 0.09, and L is the turbulent length scale, taking 0.07 of the pipe diameter. In step two, the data of the actual production site are processed to select variables and outliers, and variables related to hydrogen flow are obtained, such as hydrogen flow, temperature, pressure, etc., and the selected variables are standardized, and sliding window is used to generate time series data as input data of the overall Neural ODE, and the model outputs W and θ.

[0058] II. Establishing SST k-ω turbulent flow model solving framework based on PINN

[0059] In the current part, a SST k-ω model calculation framework based on deep neural network needs to be established, and the input and output variables and the overall calculation process of the framework need to be set according to the actual shaft furnace operation. Based on the linear layer network, the NS equation calculation framework based on SST k-ω is built, the model input uses the spatial motion coordinates of hydrogen and the recorded time steps, it is assumed that each unit of the model output represents the velocity field distribution (u, v, w) and the pressure field distribution p at different positions, and the partial derivatives with respect to space coordinates and time are directly solved based on automatic differentiation technology. After obtaining the output results of the model, the overall model training loss needs to be designed according to step four, and the loss function for PINN network training is constructed based on the velocity, pressure and corresponding differential components output by the linear layer neural network. The overall loss function contains six items, which are initial condition loss, boundary condition loss, momentum loss term, mass loss term and k and ω loss term. For the initial condition loss, it directly calculates the difference between the velocity distribution of the initial flow field and the velocity distribution of the model output flow field, and the specific formula is as follows.

[0060] L init =||u(x,y,z,t=0)-u i (x,y,z)| 2(7); after ensuring that the initial conditions meet the actual operation, the boundary conditions are very important to ensure the reasonableness of the results of the model network. For the boundary condition loss, it is composed of two parts, which are the Dirichlet boundary condition loss term and the Neumann boundary condition loss term, wherein the Dirichlet boundary condition loss term is expressed as:

[0061] L BC-Dirichlet =||u(x,y,z,t)-u b (x,y,z,t)| 2 (8);

[0062] wherein u(x,y,z,t) is the velocity field distribution of different positions at t time point output by the model, and u b (x,y,z,t) is the velocity field distribution of different positions at t time point under the actual operation. The Neumann boundary condition loss term is expressed as

[0063]

[0064] The purpose of this loss term is to ensure that the normal directional derivative of the velocity field output by the model at the boundary position is 0, that is, there is no shear flow. For the control equation part loss, it is divided into momentum equation part loss and mass equation part loss, which are as follows:

[0065]

[0066] At the same time, in order to limit the model, the limiting term of SST k-ω equation is introduced, which is shown in the following formula:

[0067]

[0068] L k =||R k | 2 (14);

[0069] L ω =||R ω | 2 (15); finally, the training loss of the whole PINN is

[0070] L=L init +L BC-Neumann +L BC-Dirichlet +L momentum +L continuity +L k +L ω (16).

[0071] After obtaining the loss, the total loss function is minimized by the Adam optimization algorithm, and the network weights are updated by backpropagation to gradually reduce the loss value, as shown in step five. During the training process, the downward trend of the total loss and its components needs to be monitored regularly to ensure that the model prediction satisfies the physical constraints and boundary conditions. The trained PINN network directly outputs the velocity distribution of the flow field, and the actual collected data is used as the label. Gradient backpropagation is performed based on the overall loss described above. Finally, when modeling the actual flow field, the trained Neural ODE and PINN network directly outputs the velocity distribution of the flow field, and only the parameters W and θ output by the Neural ODE are used. The above parameters are brought into the SST k-ω turbulence model, and the above equations are used to model the gas motion.

[0072] III. Solid particle modeling and coupling modeling based on drag force model

[0073] Based on the velocity field and pressure field distribution output by PINN, the solid phase motion is modeled by DEM, and the gas-solid motion is combined based on the drag force model. The discrete element method independently analyzes the force on each particle and models the particle motion equation based on Newton's equation. The specific force and motion equation are as follows:

[0074]

[0075] where F gp is the pressure gradient force, F drag is the drag force, F c is the collision pressure between particles, which can be divided into normal force and tangential force,

[0076] represents the tangential contact force between particle i and particle j. After considering the drag force in the solid phase motion modeling equation, the drag force term needs to be considered in the gas phase motion. This part of the coupling affects the turbulent kinetic energy and turbulent frequency, and the specific formula results are as follows

[0077]

[0078] where is the gradient of the gas velocity field, F drag is the drag force, U rel = U g -v p , represents the relative velocity of the gas phase and the solid phase γ is a model constant related to the drag force, U rel = U g -v p , C d : drag force coefficient, related to Reynolds number. By introducing the drag force into the turbulence equation, the distribution of k and ω is corrected, and the updated parameters are used to correct the velocity field distribution, and finally the gas-solid coupling motion modeling is realized.

[0079] The technical features of the above-described embodiments can be further combined. For the sake of brevity, not all possible combinations of the technical features in the above-described embodiments are described, however, as long as the combinations of the technical features do not contradict each other, they shall be considered within the scope of the present specification.

[0080] The above-described embodiments only express several implementation manners of the present application, which are described in a more specific and detailed manner, but should not be understood as a limitation on the scope of the present application. It should be noted that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present application, which shall be within the protection scope of the present application. The protection scope of the present application is given by the appended claims and any equivalent technical solutions thereof.

Claims

1. A hydrogen shaft furnace gas-solid coupling modeling method based on physical information neural network enhancement, characterized in that, The steps are: Step one: for the mixing function of SST k-omega model, use the dynamic adjustment mechanism based on neural network to convert the parameters of the mixing function into a differentiable Neural-ODE network form; Step two: from a large number of collected pure hydrogen shaft furnace data related to the running state of the furnace, extract and normalize the variables, and input these pretreated data into the Neural ODE network to output the core variables W and theta of the enhanced mixing function; Step three: establish a deep neural network based SST k-omega model calculation framework, build a network to predict the hydrogen velocity field and pressure field, and take the turbulent kinetic energy k and turbulent frequency omega as key variables related to space position and time; Step four: use the velocity field and pressure field output by the network, and the corresponding differential relationship to build a total loss function including initial condition loss, boundary condition loss and control equation residual loss; Step five: minimize the total loss function by Adam optimization algorithm to iteratively update the network parameters to reduce the error; Step six: after optimization, use the trained PINN network to directly propagate the fluid flow results, and output the final velocity and pressure field distribution based on the actual collected data as the benchmark; Step seven: based on the velocity and pressure field distribution output by the PINN network, further combine the solid phase modeling, use the discrete element method to calculate the force between solid particles, and based on Newton's law, establish the motion equation of the solid particles, and through the drag force model, couple the motion equations of the gas phase and the solid phase to realize the gas-solid coupling solution; In step one, the SST k-omega model mixing function is enhanced, first the mixing function is defined and parameterized, and the over-parameterized SST k-omega model fluid motion control equation is as follows: where p and u represent the density and velocity components of the fluid, respectively, k and ω represent the turbulent kinetic energy and the turbulent frequency, respectively, P k represents the generation term of the turbulent kinetic energy, and indicates the turbulent kinetic energy generated by the shear action of the average flow field, F1 is used as a blending function after enhancement to perform smooth transition between the k-ω and k-ε models, and the function expression is as follows: In the above formula, α, β, σ ω are empirical constants, which can be fine-tuned according to the situation, y is the normal distance between the calculation area and the wall surface, v t is the kinematic viscosity, σ k and σ ω are the turbulent viscosity diffusion coefficients, W and θ are the enhancement coefficients, which enhance the dynamic fitting ability of the mixing function, and the parameterized mixing function is fitted using the Neural ODE.

2. The method of claim 1, wherein, The equation for representing F1 mixing function using Neural ODE is as follows: Where f represents the nonlinear function of the differential term, in this method, a linear layer network is used, and the activation function in the linear network layer uses Tanh function; the input part of the Neural ODE model here is the turbulent kinetic energy, turbulent frequency, wall distance and the parameters involved in the above mixing function equation, the enhanced parameters W, theta are brought into the mixing equation to obtain the parameterized mixing function, and finally the parameterized calculation result of F1 is obtained for subsequent model calculation; the data that the shaft furnace can record includes hydrogen flow, temperature and pressure, before the Neural ODE model calculation, k, omega are calculated according to the hydrogen flow, temperature, where I = 0.16Re -1 / 8 , Re is the Reynolds number, C μ is a constant taken as 0.09, and L is the turbulent length scale taken as 0.07 of the pipe diameter; in the second step, the data of the actual production site are processed to select variables and outliers, to obtain variables related to hydrogen flow, including hydrogen flow, temperature, and pressure, and the selected variables are standardized, a sliding window is used to generate time series data as input data of the overall Neural ODE, and the model outputs W and θ.

3. The method of claim 2, wherein, In step three, the deep neural network based SST k-omega model calculation framework is established, and the input and output variables and the overall calculation process of the whole framework are set according to the actual shaft furnace running situation; Based on the linear layer network, the SST k-omega based NS equation calculation framework is built, the model input uses the spatial motion coordinates of hydrogen and the recorded time steps, it is assumed that each unit of the model output represents the velocity field distribution (u, v, w) and pressure field distribution p at different positions, and the partial derivatives with respect to space coordinates and time are directly solved based on automatic differentiation technology; After obtaining the output results of the model, the overall model training loss is designed according to step four, and the loss function for PINN network training is constructed based on the speed, pressure and corresponding differential components output by the linear layer neural network; the overall loss function contains six items, which are initial condition loss, boundary condition loss, momentum loss term, mass loss term and k and ω loss term; for the initial condition loss, the difference between the velocity distribution of the initial flow field and the velocity distribution of the model output flow field is directly calculated, and the formula is as follows L init = ||u(x, y, z, t = 0) - u i (x, y, z) || 2 (7); For the boundary condition loss, it is composed of two parts, which are Dirichlet boundary condition loss term and Neumann boundary condition loss term, wherein the Dirichlet boundary condition loss term is represented as: L Bc-Dirichlet = ||u(x, y, z, t) - u b (x, y, z, t) || 2 (8); where u(x, y, z, t) is the model output velocity field distribution at different locations at time t, u b (x, y, z, t) is the actual running velocity field distribution at different locations at time t; the Neumann boundary condition loss term is expressed as The purpose of this loss is to ensure that the normal directional derivative of the velocity field output by the model at the boundary position is 0, that is, there is no shear flow; the momentum loss and the mass loss are as follows: At the same time, in order to limit the model, the restriction term of SST k-ω equation is introduced, as shown in the following formula: L k =||R k || 2 (14); L ω =||R ω || 2 (15); After comprehensive, the training loss of the whole PINN is L = L init + L BC-Neumann + L BC-Dirichlet + L momentum + L continuity + L k + L ω (16); After obtaining the loss, according to step five, the total loss function is minimized by Adam optimization algorithm, the network weight is updated by back propagation, and the loss value is gradually reduced; during the training process, the downward trend of the total loss and its components is monitored regularly to ensure that the model prediction solution meets the physical constraints and boundary conditions; the trained PINN network is used to directly output the velocity distribution of the flow field, the actual collected data is used as the label, and the gradient back propagation is carried out based on the above overall loss; finally, in the actual modeling, the trained Neural ODE and PINN network are used to directly output the velocity distribution of the flow field, only the parameters W and θ output by the Neural ODE are used, and the above parameters are brought into the SST k-ω turbulent flow model to model the gas motion.

4. The method of claim 3, wherein, In step seven, based on the velocity field and pressure field distribution output by PINN, DEM is used to model the solid phase motion, and gas-solid motion is combined based on the drag force model; The discrete element method independently analyzes the force condition of each particle, and models the particle motion equation according to Newton's equation, and the force condition and motion equation are as follows: where F gp is the pressure gradient force, F drag is the drag force, F c is the collision pressure between particles, which is decomposed into normal and tangential forces, represents the tangential contact force between particle i and particle j; after considering the drag force into the solid phase motion modeling equation, the drag force term is considered into the gas phase motion, and the coupling mode of this part has an impact on the turbulent kinetic energy and turbulent frequency, respectively, and the formula results are as follows wherein is the gradient of the gas velocity field, F drag is the drag force, U rel = U g -v p , the relative velocity of the gas and solid phases γ is a model constant related to the drag force, C d : the drag coefficient, which is related to the Reynolds number; by introducing the drag force into the turbulent flow equation, the distribution of k and ω is corrected, and the updated parameters are used to correct the velocity field distribution, and finally the gas-solid coupling motion modeling is realized.

Images