A method and system for simulating and optimizing air flow organization based on CFD

Abstract

This invention relates to the field of airflow optimization technology, specifically a CFD-based method and system for airflow organization simulation and optimization design. The method includes: establishing a geometric model of the internal space of a dust collector and generating a computational grid; setting the local drag coefficient and opening ratio of the perforated plate as optimization parameters and performing initial sampling; using CFD to calculate the relative root mean square value and flow pattern fit of the initial samples, and training a surrogate model to establish a mapping relationship; using an optimization algorithm to globally optimize the surrogate model, and combining it with an active learning sampling algorithm to evaluate uncertainties and generate incremental samples; using CFD to verify the incremental samples, if the error meets the convergence condition, outputting the optimal solution; otherwise, updating the surrogate model and iterating cyclically. This invention constructs a closed-loop framework of CFD verification-surrogate evaluation-active learning sampling, effectively avoiding the technical bottleneck of high-dimensional parameter optimization easily getting trapped in local optima, and achieving efficient capture of the optimal airflow distribution solution.

Description

Technical Field

[0001] This invention relates to the field of airflow optimization technology, specifically to a CFD-based method and system for airflow organization simulation and optimization design. Background Technology

[0002] The scientific design of airflow organization inside dust collectors has become a core element in reducing energy consumption and ensuring equipment operation stability. Although traditional computational fluid dynamics (CFD) simulation methods can accurately calculate flow field distribution through physical solvers, they still face significant challenges in actual engineering optimization design.

[0003] When using computer-aided design to optimize the airflow organization of electrostatic precipitators, the core requirement is to find the optimal structure and resistance parameters (such as the opening ratio of each zone) of the inlet and outlet airflow distribution plates (perforated plates). However, existing optimization methods generally face the technical bottleneck of balancing computational cost and global optimization capability. Traditional methods often rely on CFD solutions across the entire physical domain during the design phase. When faced with large-scale, high-dimensional airflow distribution parameter optimization, the massive iteration of discrete CFD cases leads to an exponential increase in computation time. Furthermore, if only a small amount of discrete static CFD snapshot data is used for manual trial and error or simple interpolation, there is a lack of effective characterization of the evolution characteristics of complex nonlinear physical fields, making it easy to get trapped in local optima and difficult to efficiently and accurately output a comprehensive optimal design scheme that balances reducing backflow losses and smoothing the airflow pattern.

[0004] To address this, a CFD-based method and system for airflow organization simulation and optimization design is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a CFD-based method and system for airflow organization simulation and optimization design to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A CFD-based method for airflow organization simulation and optimization design includes: Establish a geometric model of the target space inside the electrostatic precipitator and generate a computational grid; set optimization parameter variables, including at least the local resistance coefficient and opening ratio of each zone of the inlet and outlet airflow distribution plate; perform initial sampling on the optimization parameter variables to generate an initial set of operating condition parameters, and load them as boundary conditions into the computational grid; The initial working condition parameter set is solved using a CFD solver, velocity field data is extracted, and the relative root mean square value of the airflow velocity distribution at the target section and the fit degree of the oblique airflow pattern are calculated as the target response value to construct an initial sample set. Based on the initial sample set, a surrogate model is trained to establish the mapping relationship between the optimization parameter variables and the target response value. The optimization algorithm is used to globally optimize the surrogate model, with the goal of co-optimizing the relative root mean square value of the airflow velocity distribution in the target section and the fit of the oblique airflow pattern, to generate candidate design schemes; the uncertainty of the candidate design schemes is evaluated based on the active learning sampling algorithm to generate incremental samples; The CFD is used to solve for the true target response value of the incremental sample and compared with the predicted value for verification. If the error meets the convergence condition, the optimal airflow distribution design scheme is output; otherwise, the surrogate model is updated with the incremental sample and the global optimization is returned.

[0007] Preferably, the target space includes inlet and outlet smoke boxes and an electric field area. The geometric model is divided into five sub-regions along the airflow direction: inlet flue, inlet horn, dust removal electric field area, outlet horn, and outlet flue. The inlet flue, electric field, and outlet flue are structured using hexahedral meshes similar to the airflow streamlines of the target space, while the inlet horn and outlet horn are unstructured using tetrahedral meshes. The sub-regions are merged using unstructured mesh generation technology.

[0008] Preferably, the computational grid is defined by setting the area of ​​the upper and lower boundaries of the electrostatic precipitator, including the ash hopper baffle and the top cathode hanger, as an imaginary solid wall boundary, and generating a computational grid for this fluid domain; In the computational grid, the inlet and outlet airflow distribution plates are defined as the porous medium controlled surface, and multiple independent partitions are divided according to their physical locations. The local drag coefficient and aperture ratio of each independent partition are set as the optimization parameter variables, and the local drag effect is transformed into an additional dynamic source term in the standard fluid dynamics equations, which is jointly determined by the actual drag coefficient and the apparent velocity of the local airflow in the porous medium region. The optimization parameter variables are sampled to generate an initial operating condition parameter set, and the parameters in the parameter set are transformed into the dynamic source term coefficients of the porous medium controlled surface of each independent partition, which are then loaded into the computational grid as internal boundary conditions.

[0009] Preferably, the CFD solver is a three-dimensional steady-state fluid solver that uses the pressure correction method to perform pressure and velocity coupled calculations; During the iterative solution process, the CFD solver uses the standard turbulent kinetic energy and dissipation rate two-equation model to close the Reynolds time-averaged equations, and combines the wall function method to process the calculation of turbulent physical quantities near the hypothetical solid wall boundary. To address the additional dynamic source term introduced by the controlled surface of the porous medium within the mesh, the CFD solver employs a local linearization strategy, representing the source term as a linear function of the unknown within a small range of variation of the unknown, thereby suppressing numerical diffusion and oscillations generated when the airflow penetrates the porous plate. The iteration termination convergence criterion of the CFD solver is set as follows: the dimensionless residuals of all governing equations are less than or equal to 1 × 10⁻⁶. -4 .

[0010] Preferably, the surrogate model is a Kriging surrogate model based on Gaussian process regression; The Kriging proxy model takes the local resistance coefficient and orifice ratio of each independent zone on the three airflow distribution plates at the inlet and one airflow distribution plate at the outlet of the electrostatic precipitator as input variables, and the relative root mean square value of the target cross section and the oblique airflow pattern fit as output response values ​​to construct a nonlinear mapping relationship between high-dimensional spatial parameters and airflow organization performance. When evaluating the uncertainty of candidate design schemes, the active learning sampling algorithm uses the prediction mean square error synchronously output by the Kriging surrogate model when predicting the target response value as a quantitative indicator, and prioritizes extracting the parameter combination points with the prediction flow pattern closest to the smooth oblique airflow and the largest prediction mean square error as the incremental samples.

[0011] Preferably, the active learning sampling algorithm is a composite acquisition function optimization algorithm based on the weighted sum of expected increment and prediction mean square error; the composite acquisition function is composed of a weighted combination of mining and exploration terms, wherein the mining term is the predicted relative root mean square value and the oblique airflow pattern fit output by the Kriging surrogate model, and the exploration term is the prediction mean square error output by the Kriging surrogate model; during iterative updates, the active learning sampling algorithm finds the maximum value of the composite acquisition function in the candidate design schemes, and uses the parameter combination point corresponding to the maximum value as the incremental sample.

[0012] Preferably, the optimization algorithm is a multi-objective genetic algorithm with a non-dominated sorting mechanism; The multi-objective genetic algorithm encodes the local drag coefficient and orifice ratio of each independent zone of the airflow distribution plate into chromosomes, and uses the relative root mean square value of the target section predicted by the surrogate model and the fit of the oblique airflow pattern as the fitness evaluation index. The multi-objective genetic algorithm generates a parameter population through selection, crossover, and mutation operations, performs iterative evaluation and hierarchical sorting on the surrogate model, and outputs a Pareto optimal solution set consisting of multiple sets of non-dominated solutions as the candidate design scheme.

[0013] Preferably, the multi-objective genetic algorithm merges the parent and offspring parameter populations, and performs rapid non-dominated sorting on the merged population based on two fitness evaluation indicators: the relative root mean square value of the target cross-section and the fit of the oblique airflow pattern, and divides the population into multiple non-dominated frontier levels. In the selection operation, parameter individuals with higher frontier levels are preferentially retained to enter the next generation population. When the number of parameter individuals in the same frontier level exceeds the remaining number that the next generation population needs to accommodate, the crowding distance of each parameter individual in that level is calculated, and parameter individuals with smaller crowding distances are preferentially eliminated, while parameter individuals with larger crowding distances are retained to enter the next generation population.

[0014] A CFD-based airflow organization simulation and optimization design system includes: Operating condition construction module: Establish a geometric model of the target space inside the electrostatic precipitator and generate a computational grid; set optimization parameter variables, including at least the local resistance coefficient and opening ratio of each zone of the inlet and outlet airflow distribution plate; perform initial sampling on the optimization parameter variables to generate an initial operating condition parameter set, and load it into the computational grid as boundary conditions; Sample Solving Module: The initial working condition parameter set is solved using a CFD solver, velocity field data is extracted, and the relative root mean square value of the airflow velocity distribution at the target section and the fit degree of the oblique airflow pattern are calculated as the target response value to construct an initial sample set; a surrogate model is trained based on the initial sample set to establish the mapping relationship between the optimization parameter variables and the target response value; Model optimization module: Utilizes optimization algorithms to globally optimize the surrogate model, aiming to coordinate the relative root mean square value of the airflow velocity distribution at the target cross section with the fit of the oblique airflow pattern, and generates candidate design schemes; Based on the active learning sampling algorithm, evaluates the uncertainty of the candidate design schemes and generates incremental samples; Verification Iteration Module: Utilizes CFD to solve for the true target response value of the incremental sample and compares it with the predicted value for verification. If the error meets the convergence condition, the optimal airflow distribution design scheme is output; otherwise, the surrogate model is updated with the incremental sample, and the global optimization is returned.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This application establishes a mapping relationship between optimization parameter variables (such as local drag coefficient and orifice ratio) and target response values ​​by constructing a surrogate model based on Gaussian process regression. In the global optimization stage, this application directly uses the surrogate model to replace the traditional full-physical domain CFD solution process, fundamentally avoiding the problem of exponential growth in computation time caused by massive discrete working condition iterations when optimizing high-dimensional airflow distribution parameters, and achieving a balance between computational cost control and global optimization scope.

[0016] 2. This application integrates a multi-objective genetic algorithm with a non-dominated sorting mechanism and an active learning sampling algorithm. By simultaneously considering the predicted value and the prediction mean square error through a composite acquisition function, it breaks through the limitations of traditional techniques that rely on a small number of static CFD snapshots for manual trial and error or simple interpolation. It can fully explore the evolution characteristics of the flow field in a huge design space, efficiently lock the Pareto optimal solution set that takes into account both reducing the relative root mean square value and fitting the oblique airflow pattern, and thus output a highly scientific and accurate comprehensive optimal design scheme for airflow distribution.

[0017] 3. This application focuses on the core component of the airflow distribution plate inside the dust collector. It not only optimizes the geometric space representation through a hybrid mesh generation technique that closely approximates real streamlines, but also transforms the perforated plate resistance into an additional dynamic source term. During the CFD iterative solution process, this application employs a local linearization strategy for the additional dynamic source term, suppressing the numerical diffusion and oscillation problems that are easily caused by airflow penetrating the perforated plate region. This ensures the high stability of the flow field calculation convergence process and the reliability of the final output results. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of a CFD-based airflow organization simulation and optimization design method. Figure 2 This is a schematic diagram of a system architecture for CFD-based airflow organization simulation and optimization design. Figure 3 A schematic diagram of the process for determining incremental samples. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This invention provides a CFD-based method and system for airflow organization simulation and optimization design, the technical solution of which is as follows: Please see Figure 1 A CFD-based method for airflow organization simulation and optimization design includes: Establish a geometric model of the target space inside the electrostatic precipitator and generate a computational grid; set optimization parameter variables, including at least the local resistance coefficient and opening ratio of each zone of the inlet and outlet airflow distribution plate; perform initial sampling on the optimization parameter variables to generate an initial set of operating condition parameters, and load them as boundary conditions into the computational grid; The initial working condition parameter set is solved using a CFD solver, velocity field data is extracted, and the relative root mean square value of the airflow velocity distribution at the target section and the fit degree of the oblique airflow pattern are calculated as the target response value to construct an initial sample set. Based on the initial sample set, a surrogate model is trained to establish the mapping relationship between the optimization parameter variables and the target response value. The optimization algorithm is used to globally optimize the surrogate model, with the goal of co-optimizing the relative root mean square value of the airflow velocity distribution in the target section and the fit of the oblique airflow pattern, to generate candidate design schemes; the uncertainty of the candidate design schemes is evaluated based on the active learning sampling algorithm to generate incremental samples; The CFD is used to solve for the true target response value of the incremental sample and compared with the predicted value for verification. If the error meets the convergence condition, the optimal airflow distribution design scheme is output; otherwise, the surrogate model is updated with the incremental sample and the global optimization is returned.

[0021] Example 1: A geometric model of the target space inside the electrostatic precipitator is established and a computational grid is generated. The geometric model is a closed three-dimensional fluid region that is completely connected inside, constructed by extracting the internal cavity boundary of the actual physical equipment of the electrostatic precipitator. It is used to strictly limit the physical boundary and computational range of the airflow organization simulation. The origin of the global three-dimensional space coordinate of the geometric model is preset to be located at the geometric center of the fluid inlet interface at the foremost end, and the horizontal flow direction of the dominant airflow is preset as the positive direction of the major axis of the coordinate system.

[0022] The target space includes inlet and outlet smoke boxes and an electric field area. The geometric model is divided into five sub-regions along the airflow direction: inlet flue, inlet horn, dust removal electric field area, outlet horn, and outlet flue. The inlet flue, electric field, and outlet flue are structured using hexahedral meshes similar to the airflow streamlines of the target space, while the inlet horn and outlet horn are unstructured using tetrahedral meshes. The sub-regions are merged using unstructured mesh generation technology.

[0023] Specifically, along the airflow direction, the geometric model is physically divided using 3D modeling software into five sub-regions: inlet flue, inlet horn, electric field region, outlet horn, and outlet flue. In this embodiment, the inlet flue is measured to be a cylinder with a diameter of 2 meters and a length of 3 meters; the inlet horn (i.e., inlet smoke box) is measured to be 4 meters long, with its geometric cross-section smoothly expanding from a 2-meter diameter circle to a 10-meter × 10-meter rectangle; the electric field section (i.e., electric field region) is measured to be a rectangular channel with a length of 16 meters, a width of 10 meters, and a height of 10 meters; the outlet horn (i.e., outlet smoke box) is measured to be 4 meters long, with its geometric cross-section smoothly shrinking from a 10-meter × 10-meter rectangle to a 2-meter diameter circle; and the outlet horn is measured to be a cylinder with a diameter of 2 meters and a length of 3 meters. For the irregular flow field with bends in the inlet and outlet flues, a multi-block topological mapping method is used to generate a hexahedral mesh. The radial mesh has 40 nodes, the circumferential mesh has 100 nodes, and the axial mesh size is 0.1 meters. The mesh lines are forced to align with the streamlines within the duct. For the horizontal linear motion characteristics of the airflow in the electric field section, the Cartesian orthogonal meshing method is used to generate a hexahedral mesh. The base mesh size for the horizontal, vertical, and flow directions is preset to 0.1 meters. To accurately capture the boundary layer flow near the wall, local mesh refinement is applied in the near-wall region of the electric field section near the electrodes. The first mesh layer height is preset to 0.005 meters, the number of mesh layers is set to 5, and the mesh growth rate is set to 1.2. These preset parameters ensure that the node distribution of the structured mesh matches the dominant horizontal airflow streamlines within the electric field region. Due to the abrupt changes in cross-section and variable diameter structure of the inlet and outlet horns, the internal airflow exhibits multi-directional diffusion, contraction, and oblique recirculation. The preset global maximum tetrahedral mesh size is 0.2 meters, and the minimum mesh size is 0.02 meters. An unstructured mesh generation algorithm based on the frontal advancement method is employed. First, triangular face meshes are generated on the outer geometric boundary of the horn, and then tetrahedral body meshes are grown into the internal volume of the model. Simultaneously, at the preset installation positions of the airflow distribution plate and at the geometric abrupt corners, curvature and proximity size functions are applied to automatically refine the local mesh. The preset curvature normal angle is 18°, allowing the tetrahedral mesh, without specific directional constraints, to adapt to the complex oblique and vortex airflow evolution regions within the horn.

[0024] The boundary surfaces of adjacent sub-regions are extracted as mesh stitching boundaries, specifically the four geometrically overlapping surfaces between the inlet flue and the inlet horn, the inlet horn and the electric field region, the electric field region and the outlet horn, and the outlet horn and the outlet flue. A node tolerance merging algorithm is initiated, with a preset tolerance value of 0.0001 meters for node overlap search. Adjacent mesh nodes on the interface with a distance less than this tolerance value are scanned and automatically stitched together. For misaligned regions where the mesh topology is inconsistent due to the boundary between structured and unstructured meshes, an interface interpolation method is used to transform them into non-conformal data transfer surfaces. A bidirectional interpolation logic is constructed to ensure flux conservation, and the five independently generated sub-region meshes are merged.

[0025] By modularizing and dividing the physical space into regions, a reasonable allocation of computing resources is achieved. In regular regions such as pipe sections and electric field sections, hexahedral structured meshes are used to suppress numerical pseudo-diffusion and closely match the airflow streamlines. Meanwhile, in the inlet and outlet end caps where there are abrupt changes in cross-section, tetrahedral unstructured meshes are used to adapt to complex boundaries and restore the multi-directional diffusion and eddy evolution process. Finally, the unstructured mesh merging and bidirectional interpolation technology at the interface are used to seamlessly integrate the sub-regions and ensure global flux conservation.

[0026] Set optimization parameter variables, including at least the local drag coefficient and opening ratio of each zone of the inlet and outlet airflow distribution plate; perform initial sampling on the optimization parameter variables to generate an initial operating condition parameter set, and load it into the computational grid as boundary conditions; The computational grid is defined by setting the area of ​​the upper and lower boundaries of the electrostatic precipitator, including the ash hopper baffle and the top cathode hanger, as an imaginary solid wall boundary, and generating a computational grid for this fluid domain; In the computational grid, the inlet and outlet airflow distribution plates are defined as the porous medium controlled surface, and are divided into multiple independent zones according to their physical locations. The local drag coefficient and porosity of each independent zone are set as the optimization parameter variables. The drag effect brought by the local drag coefficient is transformed into an additional dynamic source term in the standard fluid dynamics equations, which is jointly determined by the actual drag coefficient and the apparent velocity of the local airflow in the porous medium region. The optimization parameter variables are sampled to generate an initial operating condition parameter set, and the parameters in the parameter set are transformed into the dynamic source term coefficients of the porous medium controlled surface of each independent zone, which are then loaded into the computational grid as internal boundary conditions. Specifically, the complete geometric dimensions of the target space inside the dust collector are obtained, and the regions containing the ash hopper baffle and the top cathode hanger on the upper and lower boundaries of the electrostatic precipitator are identified. To avoid the surge in mesh count and computational divergence caused by performing full-detail meshing on the aforementioned complex non-mainstream regions, this embodiment sets the regions containing the ash hopper baffle and the top cathode hanger on the upper and lower boundaries of the electrostatic precipitator as hypothetical solid wall boundaries. That is, the originally uneven ash hopper baffle and top cathode hanger structures are simplified into flat or stepped continuous physical boundaries, and a computational mesh is generated for the fluid domain jointly closed by the hypothetical solid wall boundary and other conventional boundaries. The hypothetical solid wall boundary is set as a no-slip wall condition in the solver. In the generated mesh model, instead of establishing a geometric model of the solid holes on the airflow distribution plate, the spatial position corresponding to the inlet and outlet airflow distribution plates is defined as the porous medium controlled surface. In this embodiment, to avoid computational divergence caused by an excessively large mesh aspect ratio, the porous medium controlled surface is set as a thin mesh region with a numerical equivalent thickness of 50 mm in the computational domain to equivalently replace the actual physical thickness of the solid porous plate; multiple independent partitions are divided according to the physical position, that is, the porous medium controlled surface where the entire inlet and outlet airflow distribution plate is located is divided into mesh surfaces along the spatial height and width directions; in this embodiment, the inlet airflow distribution plate is divided into three regions along the height direction (upper, middle, lower) and three regions along the width direction (left, middle, right), strictly dividing nine non-overlapping independent partitions on a single distribution plate, and a total of 36 non-overlapping independent partitions are formed by the four airflow distribution plates; For each of the above-described independent zones, the local resistance coefficient and the opening ratio of each independent zone are set as the optimization parameter variables. Since each independent zone has independent opening characteristics, the above 36 independent zones constitute a total of 72 mutually independent optimization parameter variables. The local resistance coefficient is determined by combining the empirical formula for calculating the local resistance of porous thin plates in fluid mechanics with the actual thickness of the distribution plate and the opening shape parameters. The opening ratio is determined by the structural bearing capacity limit of the airflow distribution plate and the engineering design standards for preventing dust accumulation and clogging. The drag effect caused by the local drag coefficient is transformed into an additional dynamic source term in the standard fluid dynamics equations, determined by both the actual drag coefficient and the apparent velocity of the local airflow within the porous medium region. The local drag coefficient (denoted as...) ) represents the aerodynamic drag characteristic parameter on an engineering macroscopic scale, and the actual drag coefficient (denoted as ) The parameters are the internal calculation parameters required for the source term equations in the porous media model of the computational fluid dynamics solver. The specific transformation logic and steps are as follows: According to aerodynamic principles, the macroscopic pressure drop loss before and after the airflow passes through the distribution plate is... The voltage drop loss is evenly distributed to a thickness of [thickness value missing]. In the controlled surface mesh of the porous medium, the momentum decay per unit volume can be obtained, i.e., the additional dynamic source term: ; By comparing the equations of the additional dynamic source term in standard fluid mechanics that include the actual drag coefficient: ; The actual drag coefficient is mathematically equivalent to the set local drag coefficient; in the formula, For gas density, The apparent velocity of the local airflow within the porous medium region. The thickness of the controlled surface of the porous medium; The optimal Latin hypercube sampling algorithm is used to perform global random spatial sampling of the optimization parameter variables within a preset value range, generating initial working condition parameter sets with different parameter combinations. Individual parameters from these sets are extracted and mapped, transforming them into dynamic source term coefficients for the controlled surfaces of the porous media in each independent partition. The local drag coefficient value of a sampled independent partition is extracted, and based on the aforementioned derived equivalent transformation relationship, the actual drag coefficient is assigned as the local drag coefficient value and substituted into the equation to calculate the specific source term drag coefficient value required as input to the porous media model for that partition. The transformed dynamic source term coefficients corresponding to each independent partition are then loaded as internal boundary conditions into the computational grid. Specifically, these values ​​are directly assigned to the internal porous boundary attribute panel of the corresponding nine independent partition grid domains, completing the initial working condition boundary loading.

[0027] By transforming complex non-mainstream regions into hypothetical solid walls to reduce the number of meshes and improve solution stability, and by using porous medium controlled surfaces to replace solid pores and performing independent partitioning to achieve spatial fine control of drag characteristics, the local drag coefficients are accurately and equivalently transformed into additional dynamic source terms, thus opening up the mapping channel between macroscopic engineering parameters and the underlying CFD calculation boundary.

[0028] The initial working condition parameter set is solved using a CFD solver, velocity field data is extracted, and the relative root mean square value of the airflow velocity distribution at the target section and the degree of fit between the oblique airflow pattern are calculated as the target response value to construct an initial sample set. The CFD solver is a three-dimensional steady-state fluid solver that uses the pressure correction method to perform pressure and velocity coupled calculations. During the iterative solution process, the CFD solver uses the standard turbulent kinetic energy and dissipation rate two-equation model to close the Reynolds time-averaged equations, and combines the wall function method to process the calculation of turbulent physical quantities near the hypothetical solid wall boundary. To address the additional dynamic source term introduced by the controlled surface of the porous medium within the mesh, the CFD solver employs a local linearization strategy, representing the source term as a linear function of the unknown within a small range of variation of the unknown, thereby suppressing numerical diffusion and oscillations generated when the airflow penetrates the porous plate. The iteration termination convergence criterion of the CFD solver is set as follows: the dimensionless residuals of all governing equations are less than or equal to 1 × 10⁻⁶. -4 .

[0029] Specifically, the CFD solver is configured as a three-dimensional steady-state fluid solver. The CFD solver reads in a computational mesh loaded with the initial set of operating parameters as internal boundary conditions and performs coupled pressure and velocity calculations using the pressure correction method. In the specific iterative solution calculation, the solver first solves the discretized momentum equation for the current governing equation to obtain an estimated velocity field. Since this estimated velocity field usually cannot strictly satisfy the continuity equation of mass conservation, an equation regarding the pressure correction value is established based on the continuity equation and solved algebraically to obtain the pressure correction field of the fluid domain. The calculated pressure correction value is then used to numerically correct the velocity and pressure fields. By continuously repeating the above cycle of velocity calculation and pressure correction in the iterative solution until a preset convergence criterion is met, the CFD solver completes the solution for the initial set of operating parameters and finally extracts the velocity field data of the target section and the entire fluid domain from the converged flow field calculation results. In the iterative solution process, to address the unknown Reynolds stress term in the Reynolds time-averaged equation, the CFD solver employs a standard turbulent kinetic energy and dissipation rate two-equation model to close the Reynolds time-averaged equation. In this embodiment, the turbulent viscosity within the flow field is calculated by simultaneously solving the turbulent kinetic energy transport equation and the dissipation rate transport equation. Simultaneously, the wall function method is used to process the calculation of turbulent physical quantities near the boundary of the hypothetical solid wall. In this embodiment, the nodes of the first layer of grid cells near the boundary of the hypothetical solid wall are presumably located within the logarithmic law region. The wall shear stress and turbulent physical quantities at this first layer of cells are directly calculated using the semi-empirical logarithmic formula of fluid mechanics, and these are fed back as boundary conditions into the master equation. This skips the direct analysis of the complex turbulence very close to the bottom layer of the wall and completes the calculation of turbulent physical quantities near the boundary of the hypothetical solid wall.

[0030] The implementation process of the CFD solver's local linearization strategy for the additional dynamic source term is as follows: For the additional dynamic source term introduced by the controlled surface of the porous medium within the mesh, since the drag source term contains a quadratic velocity term, the CFD solver employs a local linearization strategy. Specifically, the source term is represented as a linear function of the unknown quantity within a small range of variation. That is, in the current iteration step, the source term is mathematically discretely split into a constant term and a first-order linear term containing the unknown velocity component. In this embodiment, the coefficient of the first-order linear term is calculated using the known flow velocity information from the previous iteration step, and the negative value of the coefficient is directly extracted and included in the coefficient matrix on the left main diagonal of the discrete algebraic equation system. This mathematically enhances the diagonal dominance of the algebraic equation system matrix, effectively suppressing numerical diffusion and oscillations generated when airflow penetrates the porous plate. The implementation process of setting the iteration termination convergence criterion of the CFD solver is as follows: During the iterative solution of the discrete algebraic equations by the CFD solver, the residual changes of all control equations are monitored. In this embodiment, at the end of each iteration step, the solver calculates the absolute residuals caused by imbalances in all global grid control units and normalizes them to obtain the dimensionless residuals of each control equation. In the solver control panel or calculation script, the iteration termination convergence criterion of the CFD solver is set to the condition that the dimensionless residuals of all control equations are less than or equal to 1 × 10⁻⁶. -4 When the dimensionless residuals of the continuity equation, the momentum equations in all directions, and the governing equations for turbulent kinetic energy and dissipation rate are less than or equal to 1 × 10⁻⁶ during the iteration process... -4 When the current flow field calculation has reached convergence, the solver terminates the iteration and outputs the flow field data of the target space.

[0031] The relative root mean square (RMS) value of the airflow velocity distribution at the target cross-section is calculated based on the velocity component data of each grid node, serving as the first target response value for evaluating airflow uniformity. Simultaneously, the angle distribution between the velocity vector and the target mainstream direction is extracted, and the oblique airflow pattern fit is calculated, serving as the second target response value for evaluating the flow field morphology. The initial operating parameters currently substituted into the calculation are used as input features, and the corresponding RMS values ​​and oblique airflow pattern fits are paired as output labels. The initial operating parameter set is iterated through, and the above extraction and calculation process is repeated. All paired data are integrated to construct an initial sample set. By employing the pressure correction method, the standard turbulent two-equation model, and the wall function method, the steady-state flow and turbulent evolution characteristics of the complex airflow inside the dust collector are accurately reproduced in terms of physical mechanism. At the same time, the computational efficiency and accuracy of the flow field analysis in the near-wall region are effectively taken into account. For the nonlinear resistance source term of the porous medium, a local linearization processing strategy is introduced, which effectively enhances the diagonal dominance of the solution matrix and eliminates the numerical divergence and oscillation problems that are easily caused when the airflow penetrates the porous plate interface from the mathematical level.

[0032] See Figure 3 Based on the initial sample set, a proxy model is trained, and a mapping relationship between the optimization parameter variables and the target response value is established; The surrogate model is the Kriging surrogate model based on Gaussian process regression; The Kriging proxy model takes the local resistance coefficient and orifice ratio of each independent zone on the three airflow distribution plates at the inlet and one airflow distribution plate at the outlet of the electrostatic precipitator as input variables, and the relative root mean square value of the target cross section and the oblique airflow pattern fit as output response values ​​to construct a nonlinear mapping relationship between high-dimensional spatial parameters and airflow organization performance. When evaluating the uncertainty of candidate design schemes, the active learning sampling algorithm uses the prediction mean square error synchronously output by the Kriging surrogate model when predicting the target response value as a quantitative indicator, and prioritizes extracting the parameter combination points with the prediction flow pattern closest to the smooth oblique airflow and the largest prediction mean square error as the incremental samples.

[0033] Specifically, the process of establishing and training the Kriging surrogate model based on Gaussian process regression is as follows: In the data processing software or a self-written algorithm script, an initialization of the Kriging surrogate model based on Gaussian process regression is performed. In this embodiment, the three airflow distribution plates at the inlet of the electrostatic precipitator and the one airflow distribution plate at the outlet are each divided into 9 independent zones according to a preset spatial division rule, resulting in a total of 36 independent zones from the four airflow distribution plates. The local resistance coefficient and opening ratio of each of these 36 independent zones are extracted as input variables, forming a total of 72-dimensional input feature vectors. Simultaneously, the relative root mean square value of the target cross-section and the oblique airflow pattern fit degree calculated by the fluid solver based on the initial sample set are used as output response values ​​for pairing. The covariance function of the pre-defined surrogate model is a Gaussian kernel function. The maximum likelihood estimation method is used to fit the input feature vector and output response value in the initial sample set, solve the relevant hyperparameters inside the model, and construct a nonlinear mapping relationship between 72-dimensional high-dimensional spatial parameters and two airflow organization performance evaluation indicators: the relative root mean square value of the target section and the oblique airflow pattern fit. After training the Kriging surrogate model, an optimization algorithm is used to globally optimize the surrogate model, with the goal of co-optimizing the relative root mean square value of the airflow velocity distribution at the target section and the fit of the oblique airflow pattern, to generate candidate design schemes. The optimization algorithm is a multi-objective genetic algorithm with a non-dominated sorting mechanism. The multi-objective genetic algorithm encodes the local drag coefficient and orifice ratio of each independent zone of the airflow distribution plate into chromosomes, and uses the relative root mean square value of the target section predicted by the surrogate model and the fit of the oblique airflow pattern as the fitness evaluation index. The multi-objective genetic algorithm generates a parameter population through selection, crossover, and mutation operations, performs iterative evaluation and hierarchical sorting on the surrogate model, and outputs a Pareto optimal solution set consisting of multiple sets of non-dominated solutions as the candidate design scheme.

[0034] The multi-objective genetic algorithm merges the parent and offspring parameter populations. Based on two fitness evaluation indicators—the relative root mean square value of the target cross-section and the fit of the oblique airflow pattern—it performs a rapid non-dominated sorting of the merged population and divides it into multiple non-dominated frontier levels. In the selection operation, parameter individuals with higher frontier levels are preferentially retained to enter the next generation population. When the number of parameter individuals in the same frontier level exceeds the remaining number that the next generation population needs to accommodate, the crowding distance of each parameter individual in that level is calculated, and parameter individuals with smaller crowding distances are preferentially eliminated, while parameter individuals with larger crowding distances are retained to enter the next generation population.

[0035] Specifically, the initialization and chromosome encoding process of the multi-objective genetic algorithm with a non-dominated sorting mechanism is as follows: In the optimization calculation script, the logic of the multi-objective genetic algorithm with a non-dominated sorting mechanism is configured and started. In this embodiment, the algorithm encodes the local drag coefficient and aperture ratio of each independent partition of the airflow distribution plate into chromosomes. Specifically, the local drag coefficient and aperture ratio (a total of 72 numerical variables) of the aforementioned 36 independent partitions are linearly concatenated according to a predetermined physical spatial order to generate a real-number encoded chromosome consisting of 72 gene loci. This chromosome represents a complete configuration scheme of an electrostatic precipitator airflow distribution plate. In this embodiment, the initial population size is preset to 100, that is, within the preset physical value range of the local drag coefficient and the physical value range of the aperture ratio, 100 initial real-number encoded chromosomes that meet the boundary constraints are generated through random sampling to constitute the initial parameter population.

[0036] For each chromosome in the parameter population, the 72 gene values ​​it carries are decoded and restored into a high-dimensional input feature vector, which is then directly input into the previously trained Kriging surrogate model in batches. After mathematical mapping operations, the surrogate model synchronously outputs the relative root mean square value of the predicted target cross-section and the predicted oblique airflow pattern fit degree under the corresponding chromosome configuration. Since the maximum or minimum value is usually uniformly sought in multi-objective optimization logic, in this embodiment, given that a smaller relative root mean square value represents a more uniform airflow, and a larger oblique airflow pattern fit degree represents a better flow field morphology, the system is set to use the reciprocal of the relative root mean square value and the oblique airflow pattern fit degree together as fitness evaluation indicators, so that the fitness values ​​of both objectives follow the evaluation criterion of "the larger the value, the better".

[0037] In each generation of optimization iteration, a selection mechanism based on fitness comparison is first used to select chromosomes with better performance from the current parameter population as parent individuals. Specifically, in this embodiment, two individuals are randomly selected from the current population for direct fitness comparison. The individual with better fitness is retained and added to the mating pool. This process is repeated until the mating pool is full. Simulated binary crossover is performed on pairs of parent individuals in the mating pool to generate offspring. In this embodiment, the crossover probability is preset to 0.9, and the crossover distribution index is preset to 20. Simultaneously, to maintain the spatial diversity of the population and prevent the algorithm from getting trapped in local optima, polynomial mutation is performed on the gene loci of some offspring individuals. In this embodiment, the mutation probability is preset to the reciprocal of the total gene dimension, and the mutation distribution index is preset to 20. After the above operations, a new generation of offspring parameter populations with the same size of 100 are generated.

[0038] The parent populations, each with a size of 100, are merged with the previously generated offspring populations to form a mixed population of 200. This mixed population is then input into the Kriging surrogate model to perform fitness evaluation and obtain predicted response values. Subsequently, the algorithm performs a rapid non-dominated sorting of the merged population based on two fitness evaluation indices: the relative root mean square value of the target cross-section and the fit of the oblique airflow pattern. That is, the two fitness evaluation indices of individuals in the population are compared pairwise. If both indices of individual A are not inferior to those of individual B, and at least one of the indices is strictly superior to that of individual B, then A is determined to dominate B. Based on this dominance relationship, the algorithm divides the mixed population into multiple non-dominated frontier levels. Solutions not dominated by any other individual are assigned to the first frontier level, solutions dominated only by individuals in the first frontier level are assigned to the second frontier level, and so on, until all 200 individuals are traversed. In the selection process of precisely truncating and screening 100 outstanding individuals from 200 mixed individuals to form the next generation population, the algorithm prioritizes retaining individuals with higher frontier levels into the next generation population. Specifically, individuals from the first and second frontier levels are moved in their entirety to the next generation. When the number of individuals in a certain frontier level exceeds the remaining capacity of the next generation population, the crowding distance between individuals within that level is calculated. In this embodiment, the crowding distance of an individual is set as the sum of the normalized absolute distances between the two adjacent individuals in the same frontier level across two fitness evaluation index spaces. Individuals within that level are sorted according to their calculated crowding distances, and individuals with smaller crowding distances are prioritized for elimination, while those with larger crowding distances are retained for the next generation population. This ensures the uniformity of the population distribution within the target space until the next generation population is exactly 100 individuals. In this embodiment, the maximum total number of evolutionary iterations for the multi-objective genetic algorithm is preset to 200 generations. When the iterative evaluation and hierarchical sorting loop on the surrogate model reaches the preset maximum number of evolutionary generations, the algorithm automatically stops iterating. Finally, the algorithm extracts all individuals in the first non-dominated frontier layer from the final generation population. These non-dominated solutions, which do not constitute a dominant relationship among themselves in the multi-objective dimension and represent the current optimal trade-off, together constitute the Pareto optimal solution set. The system directly outputs this Pareto optimal solution set as the candidate design scheme.

[0039] By introducing a multi-objective genetic algorithm, the mutually constraining airflow uniformity and flow pattern fit are directly used as collaborative evaluation indicators, breaking the technical limitations of traditional single-objective optimization in taking into account multi-dimensional airflow organization performance. Furthermore, by merging the parent and offspring populations and performing fast non-dominated sorting, it is ensured that excellent parameter genes are not lost in the optimization process of each generation. At the same time, the crowding distance evaluation is introduced in the critical level screening, which forces the priority retention of sparsely distributed parameter individuals, effectively maintaining the distribution uniformity and evolutionary diversity of the population in the solution space from the underlying algorithm logic.

[0040] The active learning sampling algorithm is a composite acquisition function optimization algorithm based on the weighted sum of expected increment and prediction mean square error. The composite acquisition function is composed of a weighted combination of mining and exploration terms, wherein the mining term is the predicted relative root mean square value and the oblique airflow pattern fit output by the Kriging surrogate model, and the exploration term is the prediction mean square error output by the Kriging surrogate model. During iterative updates, the active learning sampling algorithm finds the maximum value of the composite acquisition function in the candidate design schemes and uses the parameter combination point corresponding to the maximum value as the incremental sample.

[0041] The active learning sampling algorithm is an optimization algorithm based on a composite acquisition function weighted by expected increment and prediction mean square error. The specific implementation process of the composite acquisition function, which is a weighted combination of mining and exploration terms, is as follows: In the optimization iteration procedure of the surrogate model, a composite acquisition function is set to evaluate the comprehensive potential of candidate design schemes. In this embodiment, the composite acquisition function is defined as a linear dynamic weighted combination of mining and exploration terms, and its mathematical expression is set as follows: ; in, The calculated composite acquisition function value, The input candidate design schemes (i.e., high-dimensional parameter combinations containing the local resistance coefficients and open area ratios of each zone); The mining item represents the expected increment; For the aforementioned exploration item; and These are the weight coefficients for the mining item and the exploration item, respectively. To match the search requirements of the optimization algorithm at different stages, this embodiment adopts a dynamic quantization allocation strategy based on the iterative process for the weight coefficients: setting... And the weight coefficient of the exploration item With the current number of active learning iterations It exhibits exponential decay, and its specific quantitative calculation formula is as follows: ; in, Substitute the initial exploration weight upper limit parameter into the calculation; This is the preset maximum total number of active learning iterations; The constant coefficients controlling the decay rate are substituted into the calculation; in each iteration, the calculation is performed. Simultaneous calculation yields the weight coefficients of the mining item. ; It is an exponential function; The extraction term represents the expected gain in physical performance at the parameter combination point. Based on the predicted relative root mean square value and the oblique flow pattern fit from the Kriging surrogate model output, the extraction term is quantitatively constructed. First, using the minimax normalization algorithm, these two predicted values ​​from the surrogate model output are mapped to a dimensionless interval of 0 to 1, obtaining the normalized relative root mean square value and the normalized flow pattern fit. The extraction term is then determined by subtracting the normalized relative root mean square value from the normalized flow pattern fit. The exploration term represents the information gain potential and cognitive blind spot of the current parameter space location. When the Kriging surrogate model predicts the target response value of the candidate design scheme, the underlying Gaussian process regression algorithm calculates the variance of the corresponding point based on the posterior probability distribution. In this embodiment, the variance data, i.e., the prediction mean square error output by the surrogate model, is directly extracted as the exploration term. In order to match the numerical magnitude with the mining term and prevent the external weights from failing, the extracted prediction mean square error is simultaneously subjected to minimax normalization to obtain the normalized prediction mean square error, i.e., the exploration term.

[0042] During iterative updates, the active learning sampling algorithm calculates the maximum value of the composite acquisition function among the candidate design schemes and uses the parameter combination point corresponding to the maximum value as the incremental sample. The implementation process is as follows: First, hundreds or thousands of candidate design schemes generated by the global optimization algorithm are used as independent variables and input into the constructed composite acquisition function mathematical model one by one. For each specific candidate design scheme, its corresponding normalized prediction value, prediction mean square error, and dynamic weight calculated in the current iteration step are substituted to calculate and output its corresponding composite acquisition function scalar value. The composite acquisition function values ​​calculated for all candidate schemes are compared and sorted in descending order in computer memory to retrieve the unique solution that makes the composite acquisition function scalar value reach the theoretical maximum value of the current iteration step. The parameter combination point corresponding to the maximum value (i.e., a complete set of input feature data containing the local resistance coefficient and opening ratio of each specific partition) is accurately extracted from the candidate pool and used as the incremental sample of this round of active learning iteration mechanism.

[0043] By introducing a Kriging surrogate model based on Gaussian process regression, a nonlinear mapping relationship between the multidimensional physical parameters of each zone of the multiple airflow distribution plates at the inlet and outlet and the airflow performance index of the target section was successfully established, realizing a quantitative mathematical characterization of the physical characteristics of high-dimensional spatial parameters. On this basis, a dynamic active learning sampling mechanism based on a composite acquisition function was proposed, which dynamically weights and combines the mining term representing the expected performance gain with the exploration term representing the blind spot of the model's cognition. From the algorithm mechanism, this approach takes into account both the in-depth mining of known high-quality areas and the extensive traversal of the unknown parameter space.

[0044] The CFD is used to solve for the true target response value of the incremental sample and compared with the predicted value for verification. If the error meets the convergence condition, the optimal airflow distribution design scheme is output; otherwise, the surrogate model is updated with the incremental sample and the global optimization is returned. The incremental samples selected by the aforementioned active learning sampling algorithm are extracted, and the specific parameters contained in the incremental samples (i.e., the local drag coefficients and porosity of each independent partition) are transformed into additional dynamic source term coefficients for the controlled surface of the porous medium. Subsequently, the updated source term coefficients are reloaded into the computational grid of the CFD solver as internal boundary conditions. The solver is started for iterative calculation. After the dimensionless residuals converge to the preset threshold, the velocity field data of the target section is extracted, and the relative root mean square value of the target section and the oblique airflow pattern fit degree corresponding to the incremental sample are recalculated according to the preset processing logic. This is used as the true target response value corresponding to the incremental sample. The system retrieves the predicted relative root mean square value and the predicted oblique airflow pattern fit of the incremental sample previously output by the surrogate model. Then, it calculates the relative error between the actual target response value and the predicted value. In this embodiment, it calculates the absolute relative error between the actual and predicted relative root mean square values, as well as the absolute relative error between the actual and predicted flow pattern fits. This embodiment sets the convergence condition for the system's global optimization as follows: the absolute relative error of the two target response values ​​of three consecutively extracted incremental samples is less than or equal to a preset error threshold. In this embodiment, this error threshold is strictly preset to 5%. If the calculated relative error and the number of consecutively qualified samples strictly meet the above-mentioned preset convergence conditions, the system determines that the current proxy model has reached extremely high global prediction accuracy in the optimal solution region, and the entire optimization process has converged to the global Pareto optimal frontier. It then terminates all active learning and global optimization iterative cycles and outputs the parameter combination represented by the incremental samples that meet the current error conditions (i.e., a set of precise fixed configurations of local resistance coefficients and orifice ratios for each inlet and outlet zone) directly as the final optimal airflow distribution design scheme. If the relative error between the calculated true target response value and the predicted value is greater than the preset 5% error threshold, or if the number of consecutive samples meeting the convergence condition is less than the preset 3, then the surrogate model is determined to have insufficient nonlinear mapping accuracy in the current high-dimensional parameter space region, requiring further correction. Specifically, the extracted incremental sample parameter features are aligned with a set of true relative root mean square values ​​and true manifold fits obtained through high-fidelity CFD solving. This data is then used to directly supplement and merge into the existing sample set as a new set of accurate sample points, forming an expanded updated sample set. The expanded updated sample set is then used to retrain and fit the surrogate model, resolving the hyperparameters within the underlying mathematical model, thereby completing the knowledge expansion and self-updating of the surrogate model. After the surrogate model update is complete, the internal instructions are returned to the optimization algorithm module. On the updated and more accurate surrogate model, the multi-objective genetic algorithm is restarted to perform iterative evaluation and hierarchical sorting, generating a new round of candidate design schemes, and continuing to extract the next round of incremental samples by calculating the predicted mean square error. By constructing a closed-loop optimization framework of CFD high-fidelity verification, rapid evaluation of the surrogate model, and active learning intelligent sampling, this method completely breaks through the technical bottleneck of traditional dust collector airflow organization optimization, which heavily relies on massive computing resources and manual trial-and-error experience. This method uses the uniformity of airflow distribution and the fit of the oblique flow pattern as collaborative optimization objectives, highly aligning with the real physical requirements of complex dust removal projects. It utilizes an active learning algorithm to dynamically extract incremental samples with high uncertainty and feed them back to update the surrogate model. This significantly reduces the number of expensive CFD calculations while mechanistically ensuring the predictive reliability and fitting quality of the surrogate model in the global optimal solution region.

[0045] Example 2: This embodiment applies a CFD-based airflow organization simulation and optimization design system to an industrial electrostatic precipitator to solve the problems of secondary dust generation caused by excessively high local wind speeds inside the precipitator and uneven airflow distribution in the electric field area. The specific operation process of each module is as follows: See Figure 2The operating condition construction module establishes a geometric model of the target space inside the electrostatic precipitator based on the equipment's 3D drawings. It divides the space into five sub-regions along the airflow direction: inlet flue, inlet horn section, electric field region, outlet horn section, and outlet flue. A hybrid structured and unstructured mesh technology is used to generate a fluid computation mesh. To avoid computational divergence, the module sets the ash hopper baffles at the upper and lower boundaries of the electrostatic precipitator and the top cathode hanger region as imaginary solid wall boundaries. Three airflow distribution plates at the inlet and one at the outlet are defined as porous medium controlled surfaces, each divided into nine independent partitions. The system extracts the local resistance coefficients and porosity of these 36 partitions as 72-dimensional optimization parameter variables. An initial sampling algorithm is used to generate initial operating condition parameter sets with different parameter combinations. The local resistance coefficients are equivalently converted into additional power source coefficients, which are then loaded into the computation mesh as internal boundary conditions. The sample solution module invokes a three-dimensional steady-state fluid solver in the background to perform high-fidelity numerical solutions to the initial operating condition parameter set. During the iterative solution process, the pressure correction method and the standard turbulent two-equation model are used to close the Reynolds time-averaged equations, and the wall function method is combined to handle the calculation of turbulent physical quantities near the hypothetical solid wall boundary. Simultaneously, local linearization is performed on the additional power source term to suppress numerical oscillations caused by airflow penetration through the perforated plate. After the dimensionless residuals of the governing equations converge, the module automatically extracts the velocity field data of the target section and calculates the relative root mean square value and the fit between the oblique airflow pattern and the target velocity field. After completing the calculation of all initial operating condition parameter sets, the module pairs the 72-dimensional parameter variables with the calculated two target response values ​​to construct the initial sample set; it initializes a Kriging surrogate model based on Gaussian process regression to establish a nonlinear mapping relationship between the high-dimensional spatial parameters of the electrostatic precipitator and the airflow organization performance. After completing the training of the surrogate model, the model optimization module initiates a multi-objective genetic algorithm with a non-dominated ranking mechanism for global optimization. The local drag coefficient and orifice ratio of each independent zone of the airflow distribution plate are encoded as real-number chromosomes. The relative root mean square value predicted by the surrogate model and the flow pattern fit are used as fitness evaluation indicators. A parameter population is generated through selection, crossover, and mutation operations. Iterative evaluation and hierarchical ranking are performed on the surrogate model, ultimately outputting a Pareto optimal solution set composed of multiple non-dominated solutions as candidate design schemes. Next, the module initiates an active learning sampling algorithm, using a constructed composite acquisition function to evaluate the candidate schemes. The composite acquisition function dynamically weights the extraction term (i.e., the predicted target response value output by the surrogate model) and the exploration term (i.e., the predicted mean square error) based on the iteration process. The module calculates the maximum value of the composite acquisition function and extracts the parameter combination point with the predicted flow pattern closest to a smooth oblique airflow and the largest predicted mean square error as incremental samples. The verification iteration module re-converts the extracted incremental samples into boundary conditions for the power source term and resubmits them to the underlying CFD solver for absolutely realistic flow field simulation calculations. After the calculation is completed, the module compares and verifies the actual target response value with the predicted value. If the system determines that the errors of the continuously extracted incremental samples all meet the preset convergence conditions, it considers that the surrogate model has extremely high prediction reliability in the optimal solution region of the electrostatic precipitator, and then terminates the loop, outputting the incremental sample whose current error meets the conditions as the optimal design scheme for the airflow distribution of the electrostatic precipitator; otherwise, the module imports the actual incremental sample into the database to update the Kriging surrogate model, and drives the model optimization module to return to global optimization, restarting the multi-objective genetic algorithm to generate a new round of candidate schemes until the airflow optimization scheme is completely converged.

[0046] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A CFD-based method for airflow organization simulation and optimization design, characterized in that, include: Establish a geometric model of the target space inside the electrostatic precipitator and generate a computational mesh; Set optimization parameter variables, including at least the local drag coefficient and opening ratio of each zone of the inlet and outlet airflow distribution plate; perform initial sampling on the optimization parameter variables to generate an initial operating condition parameter set, and load it into the computational grid as boundary conditions; The initial working condition parameter set is solved using a CFD solver, velocity field data is extracted, and the relative root mean square value of the airflow velocity distribution at the target section and the degree of fit between the oblique airflow pattern are calculated as the target response value to construct an initial sample set. The agent model is trained based on the initial sample set, and the mapping relationship between the optimization parameter variables and the target response value is established. The optimization algorithm is used to globally optimize the surrogate model, with the goal of co-optimizing the relative root mean square value of the airflow velocity distribution in the target section and the fit of the oblique airflow pattern, to generate candidate design schemes. The uncertainty of candidate design schemes is evaluated based on the active learning sampling algorithm, and incremental samples are generated. The CFD is used to solve for the true target response value of the incremental sample and compared with the predicted value for verification. If the error meets the convergence condition, the optimal airflow distribution design scheme is output; otherwise, the surrogate model is updated with the incremental sample and the global optimization is returned.

2. The CFD-based airflow organization simulation and optimization design method according to claim 1, characterized in that, The target space includes inlet and outlet smoke boxes and an electric field area. The geometric model is divided into five sub-regions along the airflow direction: inlet flue, inlet horn, dust removal electric field area, outlet horn, and outlet flue. The inlet flue, electric field, and outlet flue are structured using hexahedral meshes similar to the airflow streamlines of the target space, while the inlet horn and outlet horn are unstructured using tetrahedral meshes. The sub-regions are merged using unstructured mesh generation technology.

3. The CFD-based airflow organization simulation and optimization design method according to claim 1, characterized in that, The computational grid is defined by setting the area of ​​the upper and lower boundaries of the electrostatic precipitator, including the ash hopper baffle and the top cathode hanger, as an imaginary solid wall boundary, and generating a computational grid for this fluid domain; In the computational grid, the inlet and outlet airflow distribution plates are defined as the porous medium controlled surface, and multiple independent partitions are divided according to their physical locations. The local drag coefficient and aperture ratio of each independent partition are set as the optimization parameter variables. The drag effect brought by the local drag coefficient is transformed into an additional dynamic source term in the standard fluid dynamics equations, which is jointly determined by the actual drag coefficient and the apparent velocity of the local airflow in the porous medium region. The optimization parameter variables are sampled to generate an initial operating condition parameter set, and the parameters in the parameter set are transformed into the dynamic source term coefficients of the porous medium controlled surface of each independent partition, which are then loaded into the computational grid as internal boundary conditions.

4. The CFD-based airflow organization simulation and optimization design method according to claim 1, characterized in that, The CFD solver is a three-dimensional steady-state fluid solver that uses the pressure correction method to perform pressure and velocity coupled calculations. During the iterative solution process, the CFD solver uses the standard two-equation model of turbulent kinetic energy and dissipation rate to close the Reynolds time-averaged equations, and combines the wall function method to handle the calculation of turbulent physical quantities near the hypothetical solid wall boundary. To address the additional dynamic source term introduced by the controlled surface of the porous medium within the mesh, the CFD solver employs a local linearization strategy, representing the source term as a linear function of the unknown within a small range of variation of the unknown, thereby suppressing numerical diffusion and oscillations generated when the airflow penetrates the porous plate. The iteration termination convergence criterion of the CFD solver is set as follows: the dimensionless residuals of all governing equations are less than or equal to 1 × 10⁻⁶. -4 .

5. The CFD-based airflow organization simulation and optimization design method according to claim 1, characterized in that, The surrogate model is the Kriging surrogate model based on Gaussian process regression; The Kriging proxy model takes the local resistance coefficient and orifice ratio of each independent zone on the three airflow distribution plates at the inlet and one airflow distribution plate at the outlet of the electrostatic precipitator as input variables, and the relative root mean square value of the target cross section and the oblique airflow pattern fit as output response values ​​to construct a nonlinear mapping relationship between high-dimensional spatial parameters and airflow organization performance. When evaluating the uncertainty of candidate design schemes, the active learning sampling algorithm uses the prediction mean square error synchronously output by the Kriging surrogate model when predicting the target response value as a quantitative indicator, and prioritizes extracting the parameter combination points with the prediction flow pattern closest to the smooth oblique airflow and the largest prediction mean square error as the incremental samples.

6. The CFD-based airflow organization simulation and optimization design method according to claim 5, characterized in that, The active learning sampling algorithm is a composite acquisition function optimization algorithm based on the weighted sum of expected increment and prediction mean square error. The composite acquisition function is composed of a weighted combination of mining and exploration terms, wherein the mining term is the predicted relative root mean square value and the oblique airflow pattern fit output by the Kriging surrogate model, and the exploration term is the prediction mean square error output by the Kriging surrogate model. During iterative updates, the active learning sampling algorithm finds the maximum value of the composite acquisition function in the candidate design schemes and uses the parameter combination point corresponding to the maximum value as the incremental sample.

7. The CFD-based airflow organization simulation and optimization design method according to claim 1, characterized in that, The optimization algorithm is a multi-objective genetic algorithm with a non-dominated sorting mechanism. The multi-objective genetic algorithm encodes the local drag coefficient and orifice ratio of each independent zone of the airflow distribution plate into chromosomes, and uses the relative root mean square value of the target section predicted by the surrogate model and the fit of the oblique airflow pattern as the fitness evaluation index. The multi-objective genetic algorithm generates a parameter population through selection, crossover, and mutation operations, performs iterative evaluation and hierarchical sorting on the surrogate model, and outputs a Pareto optimal solution set consisting of multiple sets of non-dominated solutions as the candidate design scheme.

8. The CFD-based airflow organization simulation and optimization design method according to claim 7, characterized in that, The multi-objective genetic algorithm merges the parent and offspring parameter populations. Based on two fitness evaluation indicators—the relative root mean square value of the target cross-section and the fit of the oblique airflow pattern—it performs a rapid non-dominated sorting of the merged population and divides it into multiple non-dominated frontier levels. In the selection operation, parameter individuals with higher frontier levels are preferentially retained to enter the next generation population. When the number of parameter individuals in the same frontier level exceeds the remaining number that the next generation population needs to accommodate, the crowding distance of each parameter individual in that level is calculated, and parameter individuals with smaller crowding distances are preferentially eliminated, while parameter individuals with larger crowding distances are retained to enter the next generation population.

9. A CFD-based airflow organization simulation and optimization design system, characterized in that, include: Operating Condition Construction Module: Establishes a geometric model of the target space inside the electrostatic precipitator and generates a computational mesh; Set optimization parameter variables, including at least the local drag coefficient and opening ratio of each zone of the inlet and outlet airflow distribution plate; perform initial sampling on the optimization parameter variables to generate an initial operating condition parameter set, and load it into the computational grid as boundary conditions; Sample Solving Module: The initial working condition parameter set is solved using a CFD solver, velocity field data is extracted, the relative root mean square value of the airflow velocity distribution at the target section and the fit degree of the oblique airflow pattern are calculated as the target response value, and the initial sample set is constructed. The agent model is trained based on the initial sample set, and the mapping relationship between the optimization parameter variables and the target response value is established. Model optimization module: Utilizes optimization algorithms to globally optimize the surrogate model, with the goal of coordinating the relative root mean square value of the airflow velocity distribution at the target cross section and the fit of the oblique airflow pattern, and generates candidate design schemes; The uncertainty of candidate design schemes is evaluated based on the active learning sampling algorithm, and incremental samples are generated. Verification Iteration Module: Utilizes CFD to solve for the true target response value of the incremental sample and compares it with the predicted value for verification. If the error meets the convergence condition, the optimal airflow distribution design scheme is output; otherwise, the surrogate model is updated with the incremental sample, and the global optimization is returned.

Images