Method and system for simulating slow-release pollutant transport based on data-driven turbulence modeling

Abstract

The application discloses a slow-release pollutant transport simulation method and system based on data-driven turbulence modeling, and relates to the technical field of artificial intelligence. The method comprises the following steps: performing flow field simulation on a three-dimensional ideal urban street valley physical model to obtain a flow field data set; training and improving a deep learning model by using the flow field data set, so as to obtain a turbulence viscosity prediction large model; establishing a numerical simulation method of slow-release pollutant transport and diffusion, coupling the numerical simulation method with the turbulence viscosity prediction large model, and obtaining a slow-release pollutant transport prediction numerical model; constructing a physical model of a target urban street, simulating by using the slow-release pollutant transport prediction numerical model, and obtaining a simulation result. The method makes up for the lack of a dynamic behavior prediction numerical model in the slow-release pollutant transport process of micro-scale solid powder, solves the problem of weak physical interpretability of an existing deep learning-based turbulence viscosity model, and realizes high-fidelity simulation of the gas release and migration process of slow-release solid particles in a complex urban wind field.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method and system for simulating the transport of slow-release pollutants based on data-driven turbulence modeling. Background Technology

[0002] With the rapid development of artificial intelligence technology, interdisciplinary integration has become an important trend in scientific research. The deep intersection of computational fluid dynamics (CFD), artificial intelligence (AI), and air pollution control is particularly prominent, providing new technological pathways for solving complex pollutant propagation problems. Against this backdrop, the simulation and prediction of the slow-release gaseous transport behavior of pollutants from microscale solid powders has become one of the current research hotspots in environmental science and engineering.

[0003] Current simulation needs related to atmospheric pollutant propagation mainly focus on two core dimensions: pollutant convection-diffusion simulation and turbulence simulation. In terms of pollutant convection-diffusion simulation, traditional techniques use pollutant convection-diffusion models as the core tool, describing the gas mass transport process through mathematical equations. However, this model only focuses on the convection and diffusion behavior of gases, completely neglecting the propagation dynamics of microscale solid powder pollutants in the air, and failing to characterize the diffusion behavior of newly generated gases after the slow release of gas from solid powder. This results in an inability to accurately describe the complete multiphase propagation process of solid powder propagation-gas release-gas diffusion, making it difficult to meet the simulation needs of pollution distribution and dynamic behavior of microscale solid powder pollutants in slow-release scenarios. Regarding turbulence simulation, due to the complexity of turbulence phenomena during pollutant propagation, it is necessary to characterize turbulence through modeling methods. Existing mainstream turbulence models include traditional models such as the Spalart-Allmaras model (a single-equation model commonly used in computational fluid dynamics) and the k-ε model (a two-equation model commonly used in computational fluid dynamics). These models are built upon fluid mechanics theory and provide fundamental methods for simulating turbulent processes. However, due to the limitations of simplified theoretical assumptions and the differences between the multi-scale and unsteady characteristics of actual turbulence, these traditional models suffer from insufficient prediction accuracy and applicability in complex flow scenarios. While emerging deep learning-based turbulent viscosity models show potential based on data-driven approaches, current solutions focus only on data fitting effects and do not fully incorporate the fundamental physical laws of fluid mechanics, resulting in a lack of physical interpretability and difficulty in guaranteeing the reliability of simulation results. In summary, existing numerical simulation models face the dual challenges of scenario adaptability and simulation reliability when simulating the pollution distribution and dynamic behavior of slow-release gases from microscale solid powder pollutants. Summary of the Invention

[0004] This application provides a data-driven turbulence modeling-based method and system for simulating the transport of slow-release pollutants. By integrating a deep learning model with a numerical simulation framework for slow-release pollutants, and introducing a loss function with physical constraints for joint training and optimization, a new numerical simulation solution paradigm is formed. This addresses the lack of numerical models for predicting gas distribution and dynamic behavior during the slow-release process of microscale solid powder pollutants, and solves the problems of weak physical interpretability and difficulty in adapting to complex urban environments in existing deep learning-based turbulence viscosity models. It achieves high-fidelity simulation of the gas release and migration process of slow-release solid particles in complex urban wind fields.

[0005] In a first aspect, embodiments of this application provide a method for simulating the transport of slow-release pollutants based on data-driven turbulence modeling. This includes simulating the flow field of a three-dimensional ideal urban street valley physical model to obtain flow field data at different times, forming a flow field dataset; training an improved deep learning model using the flow field dataset as samples to obtain a large-scale turbulence viscosity prediction model; establishing a numerical simulation method for the transport and diffusion of slow-release pollutants and coupling it with the large-scale turbulence viscosity prediction model to obtain a numerical model for predicting the transport of slow-release pollutants; constructing a physical model of the target urban street and performing mesh generation, then using the numerical model for predicting the transport of slow-release pollutants to perform numerical simulation, obtaining the simulation results of the flow field of the target urban street.

[0006] In conjunction with the first aspect, in one possible implementation, the flow field simulation of the three-dimensional ideal urban street valley physical model to obtain flow field data at different times includes: establishing a three-dimensional ideal urban street valley physical model; meshing the three-dimensional ideal urban street valley physical model and verifying the mesh independence and solution method stability to obtain a computational domain mesh; setting numerical boundary conditions, initial conditions, and a solution algorithm, and performing flow field simulation on the computational domain mesh to obtain flow field data at different times.

[0007] In conjunction with the first aspect, in one possible implementation, training the improved deep learning model using the flow field dataset as samples includes: extracting the velocity gradient tensor of each grid cell from the flow field dataset and converting it into multidimensional velocity gradient components; inputting the multidimensional velocity gradient components into the improved deep learning model and outputting a predicted value of turbulent viscosity.

[0008] In conjunction with the first aspect, in one possible implementation, the improved deep learning model includes: a data normalization layer for normalizing the multidimensional velocity gradient components and converting them into feature vectors; a feature extraction layer for eliminating invalid inputs in the feature vectors and converting them into dimensionless physical features; and a residual hidden block for acquiring the nonlinear relationship of the dimensionless physical features and mitigating gradient decay. The residual hidden block comprises a fully connected layer, a batch normalization layer, an activation function layer, a regularization layer, and a residual connection layer connected in sequence. Specifically, the fully connected layer expands the dimension of the low-dimensional dimensionless physical features to enlarge their feature space; the batch normalization layer standardizes the expanded dimensionless physical features to suppress gradient vanishing; and the activation function layer introduces nonlinearity. The system is activated to adapt to large fluctuations in turbulent viscosity; the regularization layer is used to randomly shield some neurons to prevent overfitting of local flow field features in the 3D ideal urban street valley physical model and ensure the continuity of physical features; the residual connection layer is used to expand the dimensionless physical features of the low dimension and add the expanded dimensionless physical features to the output of the regularization layer to avoid losing core physical information; the physical constraint output layer is used to apply physical constraints to the output turbulent viscosity prediction value; the physical constraint output layer includes a physical constraint activation layer and an output layer; the physical constraint activation layer is used to apply non-negative constraints to the turbulent viscosity prediction value and amplify the activated value to prevent numerical divergence; the output layer is used to map the high-dimensional physical features output by the residual hidden block to the low-dimensional turbulent viscosity prediction value.

[0009] In conjunction with the first aspect, in one possible implementation, training the improved deep learning model using the flow field dataset as samples further includes: marking grid regions of specific local geometric areas in the computational domain grid; and training the improved deep learning model using the flow field data of the grid regions to reduce data volume and computational cost.

[0010] In conjunction with the first aspect, in one possible implementation, the loss function of the large-scale turbulent viscosity prediction model includes: ; in, ; In the formula, L represents the loss function of the large-scale turbulent viscosity prediction model. This represents the data loss function for improving deep learning models. This represents the dimensionless adaptive weighting coefficient. Let N represent the physical constraint loss function, N represent the batch size (i.e., the number of training time steps), and M represent the number of grid cells used for training at each time step. This represents the predicted turbulent viscosity of the j-th grid cell at time i, as predicted by the improved deep learning model. R represents the true value of the turbulent viscosity at the location corresponding to the j-th grid cell at time i, and R represents the physical residual. This represents the fluid velocity at the location corresponding to the j-th grid cell at the i-th time moment, obtained from the numerical simulation. Let T denote the Hamiltonian operator, and T denote the transpose matrix. The modulus representing the average inlet velocity. Indicates the characteristic length of a building.

[0011] In conjunction with the first aspect, in one possible implementation, the method for establishing a numerical simulation of the transport and diffusion of slow-release pollutants is coupled with a large-scale turbulent viscosity prediction model to obtain a numerical model for predicting the transport of slow-release pollutants. This includes: based on the law of conservation of mass, controlling the net rate of change of mass within the incompressible fluid to zero to construct a fluid mass conservation equation; based on the law of conservation of momentum, replacing the turbulent viscosity output by the turbulent model in the viscous stress term with the predicted turbulent viscosity value to obtain a fluid dynamic conservation equation; and using the ratio of the sum of the predicted turbulent viscosity value and the molecular diffusion coefficient to the Schmidt number as the diffusion coefficient of the diffusion term, constructing a solid particle tracking model based on particle number conservation. Equations are used to simulate the dispersion behavior of solid particles in complex turbulence. Based on the diffusion coefficient, a tracking equation for solid particle concentration is constructed according to the law of conservation of mass to describe the transport and change of solid particle concentration per unit volume. Based on the law of conservation of mass, a concentration tracking equation for slow-release gas is constructed according to the diffusion coefficient, the particle concentration source term, and the contact area per unit volume to describe the combined effect of slow-release gas flow and diffusion with the fluid. The fluid mass conservation equation, fluid momentum conservation equation, solid particle tracking equation, solid particle concentration tracking equation, and slow-release gas concentration tracking equation are coupled to obtain a numerical model for predicting the transport of slow-release pollutants.

[0012] In conjunction with the first aspect, in one possible implementation, constructing a physical model of the target city streets and performing mesh generation includes: extracting map data of the area where the target city is located; extracting building outlines of the target city based on the map data; generating three-dimensional geometry based on the extracted building outlines; performing texture mapping on the three-dimensional geometry to obtain a physical model with physical properties; and performing mesh generation on the physical model.

[0013] Secondly, embodiments of this application provide a data-driven turbulence modeling-based slow-release pollutant transport simulation system, comprising: a construction module for simulating the flow field of a three-dimensional ideal urban street valley physical model to obtain flow field data at different times, forming a flow field dataset; a training module for training an improved deep learning model using the flow field dataset as samples to obtain a large-scale turbulence viscosity prediction model; a coupling module for establishing a numerical simulation method for the transport and diffusion of slow-release pollutants and coupling it with the large-scale turbulence viscosity prediction model to obtain a numerical model for predicting the transport of slow-release pollutants; and a prediction module for constructing a physical model of a target urban street and performing mesh generation, then using the numerical model for predicting the transport of slow-release pollutants to perform numerical simulation to obtain the simulation results of the flow field of the target urban street.

[0014] Thirdly, embodiments of this application provide an apparatus comprising: a processor; a memory for storing processor-executable instructions; wherein, when the processor executes the executable instructions, it implements the method as described in the first aspect or any possible implementation of the first aspect.

[0015] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: This application's embodiments train an improved deep learning model using flow field datasets as samples, fully integrating the fundamental physical laws of fluid mechanics with a data-driven approach, forming a new numerical simulation paradigm. Introducing a loss function with physical constraints for joint training and optimization effectively improves the accuracy and physical consistency of turbulent viscosity prediction. By constructing an improved deep learning model, not only is high-precision simulation of complex turbulent phenomena achieved, but the reliability and stability of the model's output results are also ensured. Through a numerical model predicting the transport of slow-release pollutants, a comprehensive description of the complete multiphase propagation process—solid powder propagation, gas release, and gas diffusion—is provided, offering strong technical support for simulating pollution distribution and dynamic behavior in microscale solid powder slow-release pollutant scenarios. This also fills the gap in numerical models for predicting gas distribution and dynamic behavior in the microscale solid powder slow-release pollutant process, effectively solving the problems of weak physical interpretability and difficulty in adapting to complex urban environments in existing deep learning-based turbulent viscosity models. Furthermore, it achieves high-fidelity simulation of the slow-release gas propagation process of microscale solid powder slow-release pollutants in real, complex urban environments, providing a scientific basis for urban environmental planning, pollution control strategy formulation, and emergency response. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating a data-driven turbulence modeling-based method for simulating the transport of slow-release pollutants, provided in this application embodiment; Figure 2 A schematic diagram of the structure of the data-driven turbulence modeling-based slow-release pollutant transport simulation system provided in this application embodiment; Figure 3 Example diagram of a three-dimensional ideal urban street valley physical model provided in the embodiments of this application; Figure 4 The velocity distribution diagram at the center sectional position of the conventional numerical simulation provided in this application embodiment; Figure 5 A velocity distribution map at the center section of the numerical model for predicting the transport of slow-release pollutants provided in this application embodiment; Figure 6 A comparison of velocities at the center of the cross-section between a traditional numerical simulation and a numerical model for predicting the transport of slow-release pollutants, as provided in this application embodiment; Figure 7 This application provides a physical model of the target city streets for its embodiments; Figure 8 An example diagram of the mesh division of the physical model of the target city streets provided in the embodiments of this application; Figure 9 The solid pollutant powder particle number distribution diagram for 1000s provided in the embodiments of this application; Figure 10 The slow-release gas concentration distribution diagram for 1000s provided in the embodiments of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0019] The following description of some technologies involved in the embodiments of this application is provided to aid understanding and should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, some descriptions of well-known functions and structures are omitted in the following description.

[0020] Figure 1 This is a flowchart of a data-driven turbulence modeling-based method for simulating the transport of slow-release pollutants, as provided in this application embodiment, including steps 101 to 104. Figure 1 This is merely one execution order shown in the embodiments of this application and does not represent the only execution order of the slow-release pollutant transport simulation method based on data-driven turbulence modeling. Where the final result can be achieved, Figure 1 The steps shown can be performed in parallel or in reverse order.

[0021] Step 101: Perform flow field simulation on the three-dimensional ideal urban street valley physical model to obtain flow field data at different times, forming a flow field dataset. In this embodiment, a three-dimensional ideal urban street valley physical model is established; the three-dimensional ideal urban street valley physical model is meshed, and the mesh independence and solution method stability are verified to obtain the computational domain mesh; numerical boundary conditions, initial conditions, and solution algorithms are set, and flow field simulation is performed on the computational domain mesh to obtain flow field data at different times.

[0022] For example, a three-dimensional ideal urban street canyon physical model is established and meshed, such as... Figure 3 As shown, the entrance section of the 3D ideal urban street valley physical model has a height of 18m, the width and height of buildings are 9m and 18m respectively, and the width and length of streets are 18m and 10m respectively. A structured rectangular mesh is used for mesh generation to ensure sufficient mesh density in key areas (such as steps and areas with drastic flow field changes) to capture subtle changes in the flow field. The total number of meshes generated in this application is 1,929,600. Mesh independence verification can also be performed by comparing the flow field simulation results under different mesh densities to select the optimal mesh density that ensures computational accuracy without excessively increasing computational load. Furthermore, the stability of the solution method needs to be verified to ensure that the flow field solution does not exhibit numerical oscillations or divergence during long-term simulations.

[0023] Furthermore, numerical boundary conditions, initial conditions, and solution algorithms are set to simulate the flow field on the computational domain mesh. When setting the numerical boundary conditions, actual flow conditions are fully considered. Numerical boundary conditions are set for inlet velocity, outlet velocity, and different physical quantities at the wall. Different parameters in the numerical simulation process are defined, including but not limited to the type of algebraic equation solver, residuals, pressure-velocity coupling iterations, etc. The flow field of the transient three-dimensional ideal urban street valley model is solved using pisoFoam (a solver specifically designed for solving incompressible transient turbulence problems). Through the above operations, the flow field is simulated on the computational domain mesh, thereby accurately obtaining flow field data at different times. This flow field data covers key information such as velocity and pressure, collectively forming a flow field dataset.

[0024] For example, the inlet velocity is set to 4.7 m / s, the turbulence model is the LES large eddy model (a model of a numerical method for simulating turbulent flow), the gas is a Newtonian gas with a viscosity of 1e-5, time series result data (i.e. flow field data) are obtained at 2-second intervals, and the calculation ends at 300 seconds.

[0025] Step 102: Train an improved deep learning model using the flow field dataset as samples to obtain a large-scale model for turbulent viscosity prediction. In this embodiment, the velocity gradient tensor of each grid cell is extracted from the flow field dataset and converted into multidimensional velocity gradient components; the multidimensional velocity gradient components are input into the improved deep learning model to output the predicted turbulent viscosity value.

[0026] Specifically, the velocity gradient tensor of each grid is extracted from the flow field dataset. The velocity gradient tensor is a second-order tensor describing the spatial variation of velocity in the flow field, defined as the partial derivative matrix of the velocity vector with respect to each coordinate. The velocity gradient tensor is decomposed into multi-dimensional velocity gradient components. These velocity gradient components can comprehensively reflect the velocity changes in the flow field and meet the input requirements for improving deep learning models. The velocity gradient components include nine components of the velocity gradient: ( u / x, u / y, u / z, v / x, v / y, v / z, w / x, w / y, w / z), where u, v, and w represent the components of velocity along the x, y, and z coordinate axes, respectively.

[0027] An improved deep learning model is trained and optimized using velocity gradient components. During training, a loss function is used to measure the difference between the predicted turbulent viscosity and the actual turbulent viscosity. The parameters of the improved deep learning model are continuously adjusted using backpropagation, gradually reducing the loss function value. Simultaneously, optimization algorithms, such as stochastic gradient descent or its variants, are employed to accelerate the convergence process of the improved deep learning model and improve training efficiency. After multiple training iterations, training stops when the loss function converges or reaches a preset number of training rounds, resulting in a well-trained large-scale turbulent viscosity prediction model. This large-scale model can accurately predict the corresponding turbulent viscosity based on the input multidimensional velocity gradient components, providing crucial data support for subsequent simulations of slow-release pollutant transport.

[0028] In this embodiment, the improved deep learning model includes: a data normalization layer for normalizing multi-dimensional velocity gradient components and converting them into feature vectors; a feature extraction layer for eliminating invalid inputs in the feature vectors and converting them into dimensionless physical features; and a residual hidden block for obtaining the nonlinear relationship of the dimensionless physical features and mitigating gradient decay. The residual hidden block includes a fully connected layer, a batch normalization layer, an activation function layer, a regularization layer, and a residual connection layer connected in sequence. The fully connected layer expands the dimension of the low-dimensional dimensionless physical features to enlarge their feature space; the batch normalization layer standardizes the expanded dimensionless physical features to suppress gradient vanishing; and the activation function layer introduces nonlinear activation to adapt to... The large fluctuations in turbulent viscosity are addressed by: a regularization layer that randomly masks some neurons to prevent overfitting of local flow field features in the 3D ideal urban street valley physical model and ensure the continuity of physical features; a residual connection layer that expands the dimensionless physical features of low-dimensional objects and adds the expanded dimensionless physical features to the output of the regularization layer to avoid losing core physical information; a physical constraint output layer that applies physical constraints to the output turbulent viscosity prediction value; the physical constraint output layer includes a physical constraint activation layer and an output layer; the physical constraint activation layer applies non-negative constraints to the turbulent viscosity prediction value and amplifies the activated value to prevent numerical divergence; and the output layer maps the high-dimensional physical features output by the residual hidden block to the low-dimensional turbulent viscosity prediction value.

[0029] Specifically, the data normalization layer receives multi-dimensional velocity gradient components and transforms them into feature vectors with a uniform scale using specific normalization methods, such as min-max normalization or Z-score normalization. This eliminates the interference of dimensional differences on improving the training of the deep learning model and lays the foundation for subsequent feature extraction. For example, the multi-dimensional velocity gradient components can be 9-dimensional or 4-dimensional (ignoring the z-direction). The normalized velocity gradient components are then flattened into one-dimensional feature vectors in a fixed order (e.g., row-major order) and used as input to the extraction layer. Because the velocity gradient components vary greatly in magnitude, the data normalization layer prevents the improved deep learning model from being dominated by large-value components, ensuring a balanced contribution of each velocity gradient component to the turbulent viscosity.

[0030] The feature extraction layer is used to further analyze these feature vectors, eliminate invalid inputs, filter invalid information (such as measurement noise and redundant components), and extract the dimensionless physical features most critical for turbulent viscosity prediction, thereby enhancing the interpretability and generalization ability of the features. Specifically, the velocity gradient tensor satisfies the incompressibility condition. ,Right now u / x+ v / y+ w / z=0, which can eliminate one redundant component (e.g.) w / z), where, The velocity is represented by u, v, and w, which represent the components of the velocity along the x, y, and z coordinate axes, respectively. Furthermore, the Pearson correlation coefficient between each velocity gradient component and the turbulent viscosity can be calculated, and velocity gradient components with low Pearson correlation coefficients (such as...) can be removed. u / z). By using a feature extraction layer, the dependence on flow scales (such as Reynolds number and geometry) can be removed, enabling the improved deep learning model to learn the universal laws governing turbulent viscosity changes, rather than the local characteristics of specific operating conditions.

[0031] The residual hidden block is a core component of the improved deep learning model. It expands the dimensionality of low-dimensional, dimensionless physical features (from 6 dimensions to 64 dimensions) through a fully connected layer, increasing the dimension of the feature space to more comprehensively capture the complex relationships between features. The batch normalization layer then standardizes the expanded dimensionless physical features (mean 0, variance 1, batch size 32-64, smoothing term 10). -5The scaling and offset parameters are both 1 / 10 of their initial values ​​and are adaptively updated during training to suppress gradient vanishing and ensure the stability of the training process. The activation function layer introduces non-linear activation functions, such as ReLU (Modified Linear Unit) or Sigmoid (S-shaped activation function), to adapt to potentially large fluctuations in turbulent viscosity and improve the expressive power of the improved deep learning model. The regularization layer randomly masks some neurons (with a dropout probability of 0.2) to prevent overfitting of the improved deep learning model to the local flow field features of the 3D ideal urban street valley physical model, ensuring the continuity of dimensionless physical features and the generalization ability of the improved deep learning model. The residual connection layer expands the low-dimensional dimensionless physical features (to 64 dimensions) and adds it to the output of the regularization layer, ensuring that core physical information is not lost during the deepening of the improved deep learning model and alleviating the gradient decay problem. Alternatively, three residual connection layers can be set here, with each layer doubling the dimensionality of the layer below it, enhancing the feature abstraction capability.

[0032] Finally, the physical constraint output layer imposes physical constraints on the output turbulent viscosity prediction (e.g., max(v)). t The physical constraint activation layer first applies a non-negativity constraint to the predicted turbulent viscosity value, ensuring that the predicted value is non-negative and consistent with physical reality. Simultaneously, it amplifies the activated predicted turbulent viscosity value to prevent numerical divergence and improve the numerical stability of the improved deep learning model. The output layer then maps the dimension-expanded dimensionless physical features output from the residual hidden block to the low-dimensional predicted turbulent viscosity value, that is, mapping the 64-dimensional output features of the residual hidden block to 1-dimensional features, completing the prediction process of the entire improved deep learning model. For example, by setting the minimum predicted turbulent viscosity value to 1e-6, it ensures that the predicted turbulent viscosity value must be non-negative, while simultaneously amplifying the predicted turbulent viscosity value by a factor of 1e5 to ensure numerical stability and prevent numerical divergence caused by excessively low viscosity.

[0033] The optimizer is Adam (a deep learning optimization algorithm with an adaptive learning rate, often used to accelerate model training convergence), and the learning rate is... The weight decays to The number of training rounds is 100-200 (based on the convergence stopping of the validation set loss).

[0034] In this embodiment of the application, training an improved deep learning model using a flow field dataset as a sample further includes: marking a grid region of a specific local geometric region in the computational domain grid; and using the flow field data of the grid region to train the improved deep learning model to reduce the amount of data and computational cost.

[0035] Specifically, within the computational domain grid, localized geometric regions (such as areas near steps, flow separation zones, etc.) are marked using the topoSetDict dictionary file (a dictionary file defining how to select grid cells, faces, or points to create regions) and the topoSet command (a command-line tool used to create / modify grid regions based on the definitions in the topoSetDict dictionary) or manually specified methods. These grid regions typically exhibit complex flow characteristics and significantly impact turbulent viscosity prediction. After marking, only the flow field data (including velocity, pressure, velocity gradient, etc.) from these local grid regions is extracted as training samples for the improved deep learning model. Since the amount of data in local grid regions is much smaller than that in the entire computational domain, the amount of data required for training can be significantly reduced, thus lowering computational costs. Furthermore, training on local grid regions allows the improved deep learning model to focus more on capturing key flow features, improving the accuracy of turbulent viscosity prediction. During training, the same loss function, backpropagation algorithm, and optimization algorithm are used to iteratively optimize the improved deep learning model until the loss function converges or the preset number of training rounds is reached, ultimately yielding a large-scale turbulent viscosity prediction model.

[0036] In this embodiment of the application, the loss function of the large-scale turbulent viscosity prediction model includes: ; in, ; In the formula, L represents the loss function of the large-scale turbulent viscosity prediction model. This represents the data loss function for improving deep learning models. This represents the dimensionless adaptive weighting coefficient. Let N represent the physical constraint loss function, N represent the batch size (i.e., the number of training time steps), and M represent the number of grid cells used for training at each time step. This represents the predicted turbulent viscosity of the j-th grid cell at time i, as predicted by the improved deep learning model. R represents the true value of the turbulent viscosity at the location corresponding to the j-th grid cell at time i, and R represents the physical residual. This represents the fluid velocity at the location corresponding to the j-th grid cell at the i-th time moment, obtained from the numerical simulation. Let T denote the Hamiltonian operator, and T denote the transpose matrix. The modulus representing the average inlet velocity. Indicates the characteristic length of a building.

[0037] Furthermore, in constructing the physical residual R, this application uses the predicted turbulent viscosity of the i-th time step and the j-th grid cell output by the improved deep learning model for the turbulent viscosity term. The velocity field involved in the physical constraint loss function (i.e., This corresponds to the value of the j-th grid cell at the i-th moment obtained by the CFD solver using the fluid momentum conservation equation in numerical simulation.

[0038] Those skilled in the art should realize that the modulus of the inlet average velocity in the above formula... and the characteristic length of the building Different combinations constitute the data loss function (i.e., data items) and physical constraint loss function The normalization factor (i.e., the physical term) is used to align the dimensions of the data and physical terms, preventing the loss function from losing its physical meaning due to differences in dimensions between the data and physical terms. Specifically, this refers to the predicted turbulent viscosity value. True value of turbulent viscosity In Chinese, the unit of viscosity is m. 2 / s, from this perspective, the normalization factor of the physical term can be derived as follows: .

[0039] Dimensionless adaptive weighting coefficients This is used to balance the order-of-magnitude difference between the data items and the physical items, scaling the errors of the two items to a similar numerical range, so as to prevent the model from being dominated by one data item during training and ignoring the other.

[0040] Step 103: Establish a numerical simulation method for the transport and diffusion of slow-release pollutants, and couple it with a large-scale turbulent viscosity prediction model to obtain a numerical model for predicting the transport of slow-release pollutants. In this embodiment, based on the law of conservation of mass, the net rate of change of mass within the incompressible fluid is controlled to be zero to construct a fluid mass conservation equation; based on the law of conservation of momentum, the predicted turbulent viscosity is replaced by the turbulent viscosity output from the turbulence model in the viscous stress term to obtain a fluid dynamic conservation equation; the ratio of the sum of the predicted turbulent viscosity and the molecular diffusion coefficient to the Schmidt number is used as the diffusion coefficient of the diffusion term, and a tracking equation for solid particles is constructed based on particle number conservation to simulate the dispersion behavior of solid particles in complex turbulence; based on the diffusion coefficient, a tracking equation for solid particle concentration is constructed based on the law of conservation of mass to describe the transport and change of solid particle concentration per unit volume; based on the law of conservation of mass, a concentration tracking equation for slow-release gas is constructed based on the diffusion coefficient, particle concentration source term, and contact area per unit volume to describe the combined effect of slow-release gas flow and diffusion with the fluid; the fluid mass conservation equation, the fluid momentum conservation equation, the solid particle tracking equation, the solid particle concentration tracking equation, and the slow-release gas concentration tracking equation are coupled to obtain a numerical model for predicting the transport of slow-release pollutants.

[0041] Specifically, the fluid mass conservation equation is as follows: .

[0042] The hydrodynamic conservation equations are as follows: .

[0043] The tracking equation for solid particles is as follows: .

[0044] The equation for tracking solid particle concentration is as follows: , in, , .

[0045] The concentration tracking equation for the slow-release gas is as follows: .

[0046] In the formula, U f Indicates fluid velocity. Let represent the fluid density, t represent time, and p represent pressure. This represents the predicted turbulent viscosity, where T represents the transpose. This indicates the number of solid particles per unit volume. Indicates the molecular diffusion coefficient. Represents the Schmitt number, This represents the source term of the solid particle tracking equation. Indicates the concentration of solid particles. Represents the source term for solid particle concentration. This indicates the rate at which solid particles release gas. This represents the contact area between a unit volume of solid particles and a fluid. This represents the volume of solid particles required to generate 1 mole of gas. Indicates the concentration of the gas.

[0047] Specifically, the fluid mass conservation equation, fluid momentum conservation equation, solid particle tracking equation, solid particle concentration tracking equation, and slow-release gas concentration tracking equation are coupled. This involves discretizing these equations and using numerical methods such as the finite volume method or finite difference method to transform the continuous partial differential equations into a discrete system of algebraic equations. During discretization, the computational domain needs to be meshed, determining the node positions, volumes, and other geometric information of each mesh cell, as well as the connection relationships between adjacent mesh cells.

[0048] Furthermore, for the fluid mass conservation equation, the net rate of change of fluid mass within each grid cell is ensured to be zero by calculating the fluid mass inflow and outflow rates for each grid cell. For the fluid momentum conservation equation, the predicted turbulent viscosity is substituted into the viscous stress term to calculate the resultant force on each grid cell, thereby updating the velocity and position of the grid cells. For the solid particle tracking equation, the diffusion and migration of solid particles within each grid cell are calculated based on the diffusion coefficient and velocity gradient, updating the position and quantity of solid particles. For the solid particle concentration tracking equation, the change in solid particle concentration within each grid cell is calculated based on the results of the solid particle tracking equation, considering processes such as solid particle release and adsorption. For the slow-release gas concentration tracking equation, the change in slow-release gas concentration within each grid cell is calculated based on the diffusion coefficient, particle concentration source term, and unit volume contact area, considering processes such as gas diffusion, release, and chemical reactions. After obtaining the discretized forms of each equation, iterative methods, such as the Gauss-Seidel iterative method or the conjugate gradient method, are used to solve the algebraic equation system, obtaining numerical solutions for physical quantities such as fluid velocity, pressure, number of solid particles, concentration of solid particles, and concentration of slow-release gas for each grid cell. By continuously iterating and updating these physical quantities until the convergence condition is met, the predicted transport results of slow-release pollutants in complex turbulence are obtained.

[0049] Figures 4 to 6 The image shows the velocity distribution at the center section of a three-dimensional ideal urban street valley physical model simulated by traditional numerical simulation and a numerical model predicting the transport of slow-release pollutants, along with a comparison example. The velocity distributions of the two models are basically consistent. Figures 7 to 8 The diagram shows an example of the physical model and mesh division of a target city street provided in this application embodiment. Figures 9 to 10 The numerical model simulation of slow-release pollutant transport in the target city streets yielded a 1000s distribution map of solid pollutant powder particles and a 1000s distribution map of slow-release gas concentration. It was found that the high concentration distribution area of ​​slow-release gas was significantly different from the distribution area of ​​solid particles. This is mainly because the gas is a slow release process during particle propagation, with a lag.

[0050] Step 104: After constructing a physical model of the target city streets and dividing it into meshes, numerical simulation is performed using a slow-release pollutant transport prediction numerical model to obtain the simulation results of the flow field of the target city streets. In this embodiment, map data of the area where the target city is located is extracted, and building outlines of the target city are extracted based on the map data; three-dimensional geometry is generated based on the extracted building outlines, and texture mapping is performed on the three-dimensional geometry to obtain a physical model with physical properties; the physical model is then divided into meshes.

[0051] like Figure 7As shown, the physical model of the target city streets is first constructed using a mesh generation example. This physical model needs to accurately reflect key features such as street width, length, building layout, and boundary conditions on both sides of the street. Next, the constructed physical model is meshed. Based on the characteristics of the street flow field, a denser mesh is used near buildings and in areas with drastic flow changes, while the mesh density is appropriately relaxed in areas with relatively stable flow fields. This approach aims to improve computational efficiency while maintaining computational accuracy.

[0052] Subsequently, the initial and boundary conditions for the simulation are set. The initial conditions include the distribution of physical quantities such as initial velocity, pressure, and temperature in the street flow field; the boundary conditions cover the velocity inlet boundary at the street entrance, the pressure outlet boundary at the exit, and the no-slip boundary of the building walls on both sides of the street. The proper setting of these conditions is crucial to the accuracy of the simulation results.

[0053] After completing the above preparations, the numerical model for predicting the transport of slow-release pollutants obtained in step 103 is applied to the numerical simulation of the physical model of the target city streets. Using numerical calculation methods, coupled equations such as the fluid mass conservation equation, fluid momentum conservation equation, solid particle tracking equation, solid particle concentration tracking equation, and slow-release gas concentration tracking equation are solved to obtain numerical solutions for physical quantities such as fluid velocity, pressure, number of solid particles, solid particle concentration, and slow-release gas concentration in each grid cell of the target city streets.

[0054] During numerical simulation, once the data retention interval is reached, the simulation results of the flow field in the target city street at that moment can be obtained, including key information such as the velocity distribution, pressure distribution, and pollutant concentration distribution. This key information can intuitively reflect the transport patterns and diffusion characteristics of slow-release pollutants in real city streets, providing a scientific basis for subsequent urban environmental planning and pollution control.

[0055] This application employs a composite loss function constructed by introducing physical constraints based on the first principles of fluid dynamics into the machine learning model. This forces the improved deep learning model to adhere to fundamental physical laws while learning from the data, thereby enhancing the physical consistency of the large-scale turbulent viscosity prediction model. Simultaneously, a novel simulation method for the transport behavior of slowly released pollutants from atmospheric microscale solid powders is established and coupled with the large-scale turbulent viscosity prediction model to obtain a numerical model for predicting the transport of slowly released pollutants. This forms a new paradigm for simulating the transport process of slowly released pollutants in complex urban environments, including the dispersion behavior of solid particles and the diffusion effect of slowly released gases. This method not only considers the influence of turbulent viscosity on pollutant transport but also ensures the rationality and reliability of the simulation results through physical constraints.

[0056] While this application provides the method operation steps as described in the embodiments or flowcharts, more or fewer operation steps may be included based on conventional or non-inventive labor. The order of steps listed in this embodiment is merely one possible execution order among many and does not represent the only execution order. In actual system or client product execution, the methods shown in this embodiment or the accompanying drawings can be executed sequentially or in parallel (e.g., in a parallel processor or multi-threaded processing environment).

[0057] like Figure 2 As shown in the figure, this application embodiment also provides a slow-release pollutant transport simulation system 200 based on data-driven turbulence modeling. The system includes: a modeling module 201, a training module 202, a coupling module 203, and a prediction module 204, as detailed below.

[0058] Module 201 is used to simulate the flow field of a three-dimensional ideal urban street valley physical model, obtain flow field data at different times, and form a flow field dataset.

[0059] Training module 202 is used to train an improved deep learning model using the flow field dataset as samples, and obtain a large model for predicting turbulent viscosity.

[0060] The coupling module 203 is used to establish a numerical simulation method for the transport and diffusion of slow-release pollutants, and couples it with a large-scale turbulent viscosity prediction model to obtain a numerical model for predicting the transport of slow-release pollutants.

[0061] The prediction module 204 is used to construct a physical model of the target city streets and perform grid division. Then, it uses a slow-release pollutant transport prediction numerical model to perform numerical simulation and obtain the simulation results of the flow field of the target city streets.

[0062] Some modules of the system described in this application can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, classes, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via communication networks. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0063] The systems or modules described in the above embodiments can be implemented by computer chips or physical entities, or by products with certain functions. For ease of description, the above systems are described by dividing them into various modules based on their functions. When implementing the embodiments of this application, the functions of each module can be implemented in one or more software and / or hardware. Of course, a module that implements a certain function can also be implemented by combining multiple sub-modules or sub-units.

[0064] The methods, systems, or modules described in this application can be implemented in a computer-readable program code manner. The controller can be implemented in any suitable manner, such as a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicon Labs C8051F320. Memory controllers can also be implemented as part of the control logic of memory. Those skilled in the art will also recognize that, in addition to implementing the controller in purely computer-readable program code manner, the same functionality can be achieved by logically programming the method steps to make the controller take the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the systems included within it for implementing various functions can also be considered structures within the hardware component. Alternatively, a system used to implement various functions can be viewed as either a software module implementing the method or a structure within a hardware component.

[0065] This application also provides an apparatus, the apparatus comprising: a processor; a memory for storing processor-executable instructions; wherein, when the processor executes the executable instructions, it implements the method described in this application.

[0066] This application also provides a non-volatile computer-readable storage medium storing a computer program or instructions thereon, which, when executed, enables the method described in this application embodiment to be implemented.

[0067] Furthermore, in the various embodiments of the present invention, each functional module can be integrated into a processing module, or each module can exist independently, or two or more modules can be integrated into a single module.

[0068] The aforementioned storage media include, but are not limited to, Random Access Memory (RAM), Read-Only Memory (ROM), Cache, Hard Disk Drive (HDD), or Memory Card. The memory can be used to store computer program instructions.

[0069] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary hardware. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product, or it can be embodied in the process of data migration. The computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, mobile terminal, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0070] The various embodiments described in this specification are presented in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. All or part of this application can be used in numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, mobile communication terminals, multiprocessor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0071] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of this application.

Claims

1. A method for simulating the transport of slow-release pollutants based on data-driven turbulence modeling, characterized in that, include: Flow field simulation was performed on a three-dimensional ideal urban street valley physical model to obtain flow field data at different times, which constituted a flow field dataset. An improved deep learning model was trained using the aforementioned flow field dataset as a sample to obtain a large-scale model for predicting turbulent viscosity. A numerical simulation method for the transport and diffusion of slow-release pollutants is established and coupled with a large-scale turbulent viscosity prediction model to obtain a numerical model for predicting the transport of slow-release pollutants. This model includes: constructing a fluid mass conservation equation by controlling the net rate of change of mass within the incompressible fluid to zero based on the law of mass conservation; obtaining a fluid dynamic conservation equation by replacing the turbulent viscosity output from the turbulent model in the viscous stress term with the predicted turbulent viscosity based on the law of momentum conservation; and constructing a solid particle tracking equation based on particle number conservation, using the ratio of the sum of the predicted turbulent viscosity and the molecular diffusion coefficient to the Schmidt number as the diffusion coefficient for the diffusion term. The dispersion behavior of particles in complex turbulent flow is studied. Based on the diffusion coefficient, a tracking equation for solid particle concentration is constructed according to the law of conservation of mass to describe the transport and change of solid particle concentration per unit volume. Based on the law of conservation of mass, a concentration tracking equation for slow-release gas is constructed according to the diffusion coefficient, particle concentration source term, and contact area per unit volume to describe the comprehensive effect of slow-release gas flow and diffusion with the fluid. The fluid mass conservation equation, fluid momentum conservation equation, solid particle tracking equation, solid particle concentration tracking equation, and slow-release gas concentration tracking equation are coupled to obtain a numerical model for predicting the transport of slow-release pollutants. After constructing a physical model of the target city streets and dividing it into grids, numerical simulations are performed using the slow-release pollutant transport prediction numerical model to obtain simulation results of the flow field of the target city streets.

2. The method according to claim 1, characterized in that, The flow field simulation of the three-dimensional ideal urban street valley physical model yields flow field data at different times, including: Establish a three-dimensional ideal urban street canyon physical model; The physical model of a three-dimensional ideal urban street valley is meshed, and the mesh independence and stability of the solution method are verified to obtain the computational domain mesh. By setting numerical boundary conditions, initial conditions, and a solution algorithm, flow field simulation is performed on the computational domain grid to obtain flow field data at different times.

3. The method according to claim 1, characterized in that, The step of training an improved deep learning model using the flow field dataset as samples includes: Extract the velocity gradient tensor of each grid cell from the flow field dataset and convert it into multidimensional velocity gradient components; The multidimensional velocity gradient components are input into the improved deep learning model, which outputs a predicted value of turbulent viscosity.

4. The method according to claim 3, characterized in that, The improved deep learning model includes: A data normalization layer is used to normalize the multidimensional velocity gradient components and convert them into feature vectors. The feature extraction layer is used to eliminate invalid inputs in the feature vector and convert the feature vector into dimensionless physical features; The residual hidden block is used to obtain the nonlinear relationship of the dimensionless physical features and to alleviate gradient decay; the residual hidden block includes a fully connected layer, a batch normalization layer, an activation function layer, a regularization layer and a residual connection layer connected in sequence. The fully connected layer is used to expand the dimensionless physical features of the low dimension to broaden their feature space; the batch normalization layer is used to standardize the expanded dimensionless physical features to suppress gradient vanishing; the activation function layer is used to introduce nonlinear activation to adapt to large fluctuations in turbulent viscosity; the regularization layer is used to randomly shield some neurons to prevent overfitting of local flow field features in the three-dimensional ideal urban street valley physical model and ensure the continuity of physical features; the residual connection layer is used to expand the dimensionless physical features of the low dimension and add the expanded dimensionless physical features to the output of the regularization layer to avoid losing core physical information. A physical constraint output layer is used to apply physical constraints to the output turbulent viscosity prediction value; the physical constraint output layer includes a physical constraint activation layer and an output layer; The physical constraint activation layer is used to impose non-negative constraints on the predicted turbulent viscosity value and to amplify the activated value to prevent numerical divergence; the output layer is used to map the high-dimensional physical features output by the residual hidden block to the low-dimensional predicted turbulent viscosity value.

5. The method according to claim 2, characterized in that, The method of training the improved deep learning model using the flow field dataset as samples also includes: Mark a grid region representing a specific local geometry within the computational domain grid; The improved deep learning model is trained using the flow field data of the grid region to reduce the amount of data and computational cost.

6. The method according to claim 1, characterized in that, The loss function of the large-scale turbulent viscosity prediction model includes: ； in, ; In the formula, L represents the loss function of the large-scale turbulent viscosity prediction model. This represents the data loss function for improving deep learning models. This represents the dimensionless adaptive weighting coefficient. Let N represent the physical constraint loss function, N represent the batch size (i.e., the number of training time steps), and M represent the number of grid cells used for training at each time step. This represents the predicted turbulent viscosity of the j-th grid cell at time i, as predicted by the improved deep learning model. R represents the true value of the turbulent viscosity at the location corresponding to the j-th grid cell at time i, and R represents the physical residual. This represents the fluid velocity at the location corresponding to the j-th grid cell at the i-th time moment, obtained from the numerical simulation. Let T denote the Hamiltonian operator, and T denote the transpose matrix. The modulus representing the average inlet velocity. Indicates the characteristic length of a building.

7. The method according to claim 1, characterized in that, The process of constructing a physical model of the target city streets and dividing it into grids includes: Extract map data of the target city's region, and extract building outlines of the target city based on the map data; A three-dimensional geometry is generated based on the extracted building outline, and texture mapping is performed on the three-dimensional geometry to obtain a physical model with physical properties. The physical model is then meshed.

8. A data-driven turbulence modeling-based slow-release pollutant transport simulation system for implementing the method of any one of claims 1-7, characterized in that, include: The module is used to simulate the flow field of a three-dimensional ideal urban street valley physical model, obtain flow field data at different times, and form a flow field dataset. The training module is used to train an improved deep learning model using the flow field dataset as samples to obtain a large model for predicting turbulent viscosity. The coupling module is used to establish a numerical simulation method for the transport and diffusion of slow-release pollutants, and couples it with a large-scale turbulent viscosity prediction model to obtain a numerical model for predicting the transport of slow-release pollutants. This model includes: based on the law of mass conservation, controlling the net rate of change of mass within the incompressible fluid to zero to construct a fluid mass conservation equation; based on the law of momentum conservation, replacing the turbulent viscosity output by the turbulent model in the viscous stress term with the predicted turbulent viscosity to obtain a fluid dynamic conservation equation; and using the ratio of the sum of the predicted turbulent viscosity and the molecular diffusion coefficient to the Schmidt number as the diffusion coefficient for the diffusion term, constructing a tracking equation for solid particles based on particle number conservation to simulate... The dispersion behavior of solid particles in complex turbulent flow is studied. Based on the diffusion coefficient, a tracking equation for solid particle concentration is constructed according to the law of conservation of mass to describe the transport and change of solid particle concentration per unit volume. Based on the law of conservation of mass, a concentration tracking equation for slow-release gas is constructed according to the diffusion coefficient, particle concentration source term, and contact area per unit volume to describe the combined effect of slow-release gas flow and diffusion with the fluid. The fluid mass conservation equation, fluid momentum conservation equation, solid particle tracking equation, solid particle concentration tracking equation, and slow-release gas concentration tracking equation are coupled to obtain a numerical model for predicting the transport of slow-release pollutants. The prediction module is used to construct a physical model of the target city streets and perform grid division, and then use the slow-release pollutant transport prediction numerical model to perform numerical simulation to obtain the simulation results of the flow field of the target city streets.

9. An apparatus for performing a data-driven turbulence modeling-based method for simulating the transport of slow-release pollutants, characterized in that, include: processor; Memory used to store processor-executable instructions; When the processor executes the executable instructions, it implements the method as described in any one of claims 1 to 7.

Images