Computational grid adjustment method and device for gas compressor and storage medium

Abstract

The embodiment of the invention discloses a compressor computational grid adjustment method and device and a storage medium. According to the method, reference geometry and a reference computational grid are obtained under a reference working condition, quality parameters are extracted, and preset conditions are set; based on the transformation matrix from the reference geometry to the target working condition, constructing and training a neural network, inputting reference calculation grid data under the reference working condition, outputting a two-dimensional grid of the target working condition, and keeping topology and the number of nodes unchanged; carrying out quality verification on the output grid, and updating neural network parameters according to a verification result so as to carry out iterative output; the two-dimensional network under the target working condition after iteration is completed is stacked into a three-dimensional grid in the preset direction, on the premise that the grid topology and the number of nodes are kept unchanged, repeated grid division is reduced, the multi-working-condition grid generation efficiency and stability are improved, cross-working-condition comparison and boundary condition multiplexing are supported, the negative grid risk is reduced, and the generation efficiency of the multi-working-condition grid is improved. And meanwhile, integration with an existing simulation process is facilitated.

Description

Technical Field

[0001] This application relates to the field of simulation computing, and in particular to a method, apparatus and storage medium for adjusting the computational grid of a compressor. Background Technology

[0002] Computational Fluid Dynamics (CFD) technology, as a simulation tool, can significantly reduce the research cost in the field of fluid dynamics and is suitable for conducting extensive research on fluid dynamic mechanisms, bringing numerous conveniences to fluid mechanics research. Generally, CFD calculations mainly include model building, mesh generation, setting boundary and initial conditions, and solution calculation. Among these, mesh generation is a crucial step in CFD, and the quantity and quality of the mesh are key factors affecting the speed and accuracy of CFD solutions.

[0003] For structured meshes, in 3D CFD mesh generation, a two-dimensional mesh is first generated, followed by the superposition of multiple layers of two-dimensional meshes to create the 3D mesh. In practical engineering, the geometry of the simulation model is not static; it is often necessary to modify the model and re-mesh it for each calculation, thus increasing the overall simulation cycle. For example, when performing dynamic simulations of multi-stage compressors with adjustable guide vanes and stator vanes, the angle of the adjustable guide vanes varies under different operating conditions, requiring changes to the model and mesh, which in turn increases the simulation cycle.

[0004] Currently, while reconstructing geometry and remeshing for each simulation case can achieve high local quality, it suffers from low efficiency, long computation time, and difficulty in automation. Furthermore, the interfaces and node numbers for different simulation cases are rarely completely consistent, affecting cross-simulation comparisons. Complex channels and narrow gaps often require significant manual intervention and quality adjustments. In addition, different software and parameter settings can lead to fluctuations in mesh quality and convergence characteristics. Therefore, a meshing and adjustment method to improve simulation efficiency is urgently needed. Summary of the Invention

[0005] Embodiments of this application provide a method, apparatus, and storage medium for adjusting the computational grid of a compressor, thereby reducing the grid generation cycle during dynamic simulation of an adjustable model.

[0006] To address the aforementioned technical problems, embodiments of this application disclose the following technical solutions: In a first aspect, a method for adjusting a compressor computational grid is provided, comprising: acquiring a reference geometry of the compressor under a reference operating condition and its corresponding reference computational grid, wherein the reference computational grid includes a three-dimensional grid formed by stacking multiple two-dimensional grids under the reference operating condition along a preset direction; acquiring the quality parameters of the reference computational grid, and subjecting the reference computational grid to quality constraints based on the quality parameters and preset conditions; constructing and training a neural network model based on the reference computational grid satisfying the preset conditions and a transformation matrix for adjusting the reference geometry to a target operating condition, thereby acquiring a mapping of the two-dimensional grid from the reference operating condition to the target operating condition; inputting the mapping and the two-dimensional grid under the reference operating condition into the trained neural network model to output a two-dimensional grid under the target operating condition, while keeping the topology and the number of grid nodes unchanged; performing quality verification on the two-dimensional grid under the target operating condition based on the preset conditions, and adjusting the parameters of the neural network model according to the verification results until the two-dimensional grid under the target operating condition meets the preset conditions; and stacking the two-dimensional grid under the target operating condition output by the neural network model along the preset direction to form a three-dimensional grid under the target operating condition.

[0007] Furthermore, the compressor is a multi-stage compressor, and its reference geometry includes at least one adjustable guide vane. The reference calculation grid is generated through structured partitioning to ensure no negative grid. The preset direction is the blade height direction. The target operating condition is achieved by adjusting the angle of the at least one adjustable guide vane.

[0008] Furthermore, the quality parameters of the benchmark computation grid include at least grid orthogonality, grid aspect ratio, and grid scaling ratio. The preset condition is that all values ​​of the quality parameters meet preset thresholds, or that all values ​​of the quality parameters are optimal values ​​when the number of grid nodes does not exceed a preset limit. When the values ​​of the quality parameters do not meet the preset thresholds, the grid partitioning strategy is adjusted to iterate the quality parameters so that the values ​​of the quality parameters are closer to the preset thresholds. When the number of grid nodes exceeds the preset limit, the iteration stops, and the current values ​​of the quality parameters are taken as the optimal values.

[0009] Furthermore, the neural network model is a fully connected neural network model; the input layer of the fully connected neural network includes two-dimensional leaf shapes with different cross-sectional heights, two-dimensional computational domain range under the reference working condition, two-dimensional grid node coordinates under the reference working condition, and the transformation matrix; the output layer outputs the two-dimensional grid node coordinates or grid node displacements under the target working condition corresponding to different cross-sections, and corresponds one-to-one with the grid nodes under the reference working condition.

[0010] Furthermore, the hidden layer of the fully connected neural network uses ReLU as the activation function; the basic loss during model training is the mean squared error function, the input of which is the grid node coordinates predicted by the model and the grid node coordinates under the target working condition, and the output is the squared mean of the difference between the input quantities.

[0011] Furthermore, the quality verification and the adjustment of the parameters of the neural network model based on the verification result include: obtaining the quality parameters of the two-dimensional mesh under the target working condition, the quality parameters including at least mesh orthogonality, mesh aspect ratio, and mesh scaling ratio; comparing the quality parameters of the two-dimensional mesh under the target working condition with the preset conditions; when all values ​​of the quality parameters of the two-dimensional mesh under the target working condition meet the preset conditions, determining that the verification result is passed, and outputting the current two-dimensional mesh; when the quality parameters of any two-dimensional mesh do not meet the preset conditions, recording the excess amount of the quality parameter that does not meet the preset conditions relative to the preset conditions and the mesh position corresponding to the maximum excess amount of the ...

[0012] Furthermore, it also includes: when outputting a two-dimensional mesh under the target operating condition, performing boundary projection on the mesh nodes located at the solid wall boundary to fit the geometric boundary under the target operating condition; after forming the three-dimensional mesh under the target operating condition, docking and merging the three-dimensional mesh of the adjustable geometry in the reference geometry with the three-dimensional mesh of the fixed geometry at a predefined interface boundary to obtain a complete three-dimensional computational mesh for the compressor.

[0013] Secondly, a compressor computational grid adjustment device is provided, comprising: a reference grid acquisition module, used to acquire reference geometry and its corresponding reference computational grid under reference operating conditions, and to acquire the quality parameters of the reference computational grid for quality constraints; a mapping model training module, used to construct and train a neural network model based on the reference computational grid that meets preset conditions and a transformation matrix for adjusting the reference geometry to the target operating condition, so as to acquire the mapping of the two-dimensional grid from the reference operating condition to the target operating condition; a grid quality constraint unit, used to perform quality verification on the two-dimensional grid under the target operating condition, and record the default excess amount and the grid position corresponding to the maximum default excess amount when the preset conditions are not met; a closed-loop update control unit, used to add a quality penalty term to the model training process to form an update loss, update the parameters of the neural network model and re-output the two-dimensional grid under the target operating condition until the preset conditions are met or the number of iterations reaches a preset upper limit; and a target grid generation module, used to output the two-dimensional grid under the target operating condition while keeping the topology and the number of grid nodes unchanged, and to stack them along a preset direction to form a three-dimensional grid.

[0014] Furthermore, the neural network model is a fully connected neural network; the input layer of the fully connected neural network includes two-dimensional leaf shapes of different cross-sectional heights, the two-dimensional computational domain range under the reference working condition, the two-dimensional grid node coordinates under the reference working condition, and the transformation matrix; the output layer outputs the two-dimensional grid node coordinates or grid node displacements under the target working condition corresponding to different cross-sections, and corresponds one-to-one with the grid nodes under the reference working condition; the hidden layer of the fully connected neural network uses ReLU as the activation function, the basic loss during model training is the mean squared error function, and the update loss is formed by weighted summation with the quality penalty term provided by the grid quality constraint unit.

[0015] Thirdly, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, is used to implement any of the compressor computational grid adjustment methods provided in the first aspect above.

[0016] The above-mentioned technical solutions have the following advantages or beneficial effects: Without changing the mesh topology and the number of mesh nodes, node coordinates are directly adjusted according to the geometric transformation of the target working condition, maintaining consistency between node numbers and interface relationships, facilitating the reuse of boundary conditions and periodic boundaries, and improving the consistency and data reusability of cross-working-condition comparisons; Mesh quality constraints are introduced during the generation stage to check orthogonality, aspect ratio, scaling ratio, and positive element area. When standards are not met, the excess amount and worst position are recorded, and the model output is iteratively corrected and updated accordingly, reducing the risk of negative meshes and quality degradation, and improving computational efficiency. Stability and convergence reliability; relying on the trained neural network mapping, the target working condition 2D mesh can be generated in batches and quickly, significantly reducing reconstruction geometry and repeated meshing, and shortening the preprocessing cycle; for near-wall and narrow gap regions, the local mesh quality is maintained while keeping the topology unchanged through the coordinated control of boundary projection and quality constraints, reducing subsequent smoothing and local re-meshing; the 2D meshes of each section are stacked along the preset direction to construct a 3D mesh, and are connected and merged with the fixed geometric mesh at the predefined interface boundary to form the whole machine computing mesh. The process is easy to integrate with existing simulation systems, taking into account both engineering efficiency and feasibility. Attached Figure Description

[0017] The technical solution and its beneficial effects will become apparent from the following detailed description of specific embodiments of this application, in conjunction with the accompanying drawings.

[0018] Figure 1 Flowchart of the compressor computational grid adjustment method provided in this application; Figure 2 A flowchart of a further compressor computational grid adjustment method provided for this application; Figure 3 A schematic diagram of a compressor computational grid adjustment device 10 provided in this application; Figure 4 A schematic diagram of the adjustable guide vane provided in this application; Figure 5 The flowchart for quality verification and neural network parameter adjustment provided in this application Figure 6 A schematic diagram of the reference calculation grid for the adjustable guide vane under the reference working condition provided in this application; Figure 7 This is a schematic diagram of the neural network output provided in this application.

[0019] Explanation of reference numerals in the attached figures: 1. Guide vane under baseline operating conditions; 2. Guide vane under target operating conditions; 3. Rotation center; 10. Compressor computational grid adjustment device; 11. Baseline grid acquisition module; 12. Mapping model training module; 13. Grid quality constraint unit; 14. Closed-loop update control unit; 15. Target grid generation module. Detailed Implementation

[0020] To make the objectives, technical solutions, and beneficial effects of this application clearer, the following detailed description, in conjunction with the accompanying drawings and specific embodiments, further illustrates this application. It should be understood that the specific embodiments described in this specification are merely for explaining this application and are not intended to limit it.

[0021] In the description of this application, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," "joining," "coupling," "docking," "merging," and "overlaying" should be interpreted broadly, encompassing fixed connections, detachable connections, or integral molding, as well as mechanical / electrical / data connections; they can be direct connections or indirect connections through an intermediate medium. For meshes and geometry, "connection / docking / merging / overlaying" also includes conformal (one-to-one node correspondence) or non-conformal (continuity of quantity achieved through interpolation / matching) splicing relationships at interface boundaries. Terms such as "set at / located on / above / below / inside / outside / before / after" are used only to describe relative positions and orders, and do not constitute a limitation on absolute spatial orientation. Terms such as "first," "second," and "third" are used only to distinguish objects, and do not limit their quantity, order, or importance. Terms such as "including / contains / has" are open-ended terms, indicating the existence of, but not limited to, the listed elements; the terms "at least one / at least one" and "several" should be understood as "one or more" or "one or more". The term "and / or" indicates any or all combinations in a parallel relationship. The term "about / approximately" is to be understood, unless otherwise specified, as a value within the range of engineering tolerances permitted in the art. Unless otherwise specified, all numerical ranges include endpoints. The term "based on" should be understood as "at least partially based on".

[0022] Furthermore, in the description of this application, the engineering terms "baseline operating condition / target operating condition" refer to different operating conditions used for simulation; "baseline geometry" refers to the geometry corresponding to the baseline operating condition, and "dynamic / target geometry" refers to the geometry under other operating conditions relative to the baseline operating condition; "fixed geometry" refers to components whose shape does not change when operating conditions change, and "adjustable geometry" refers to components whose shape / attitude can change with operating conditions without changing the geometric topology (e.g., adjustable guide vanes). The term "preset direction" in compressor embodiments is usually the blade height direction (span), but does not exclude the superposition and reconstruction of other appropriate directions. Terms such as "computational grid / grid / 2D grid / 3D grid / grid cell / node / topology / number of nodes" follow the general meaning in the field of simulation computing: "unchanged topology" and "unchanged number of nodes" mean that the connectivity of the grid and the count of nodes / cells remain consistent with the baseline operating condition, and only the node coordinates are allowed to change. "Mesh quality parameters" and "quality indicators" include, but are not limited to, mesh orthogonality, mesh aspect ratio, mesh scaling ratio, and cell area positivity; "preset conditions" and "thresholds" can be fixed thresholds or constraint values ​​determined by combining measured values ​​from the benchmark mesh. "Excess amount of violation" is used to characterize the amount by which the quality indicator exceeds or falls short of the preset conditions. "Mesh location" can be identified by cell index, mesh (i, j) number, or cell centroid coordinates.

[0023] It should also be noted that the term "transformation matrix / transformation operator" describes the relative transformation from the reference geometry to the target geometry. This can be a combination of affine transformations (rotation, scaling, translation, shearing) or nonlinear transformations such as parametric / spline transformations, and can be categorized into leaf-shaped geometric transformations and 2D computational domain transformations. It can serve as conditional information for model input or be used for initial analytical transformations of the reference coordinates. "Mapping / node coordinate mapping" refers to the functional relationship between the node coordinates of the reference 2D mesh and the node coordinates of the target 2D mesh, while maintaining the same topology and number of nodes. Terms such as "neural network model / fully connected neural network / input layer / hidden layer / output layer / activation function / ReLU / training / inference / loss function / mean squared error / quality penalty term / update loss / parameter update / loop closure" have their conventional meanings in the field of machine learning. The "training" phase is used to learn the mapping relationship from input to output, and the inference phase is used to generate the target result; "basic loss" and "mean squared error" refer to the mean squared difference between the predicted node coordinates (or displacements) and the target node coordinates (or displacements); "quality penalty" applies a penalty to elements that do not meet the preset conditions based on the mesh quality index; "closed loop" and "iteration" refer to the process of recording relevant information when the quality does not meet the preset conditions and using it to update the model before outputting again; "boundary projection" refers to projecting / attaching nodes located on solid walls / interfaces onto the target geometric boundary to ensure geometric consistency.

[0024] In this application, terms such as "module / unit / device / system / server / processor / storage medium / computer program" can be implemented through software, hardware, or a combination of both; "computer-readable storage medium" can be any non-transitory storage medium. Terms such as "acquire / collect / read / receive / calculate / determine / output / generate" are used interchangeably without ambiguity, referring to routine data processing operations performed by the system / device / program. Those skilled in the art can reasonably understand the above terms based on the specific context and make equivalent substitutions for their specific implementations without departing from the spirit and scope of this application.

[0025] In the operation and design verification of multi-stage compressors, adjustable components such as adjustable guide vanes undergo attitude changes under different operating conditions, necessitating frequent multi-condition numerical simulations. Due to the complex flow channels and thin near-wall characteristics, the orthogonality, aspect ratio, scaling ratio, and element area positivity of the computational mesh significantly impact simulation stability and result reliability. Traditional methods often involve remodeling and meshing for each target operating condition. While this can achieve high local quality, it suffers from long cycles, strong reliance on manual intervention, and difficulty in achieving consistency between nodes and interfaces across operating conditions. Another approach, based on elastic bodies or RBF methods, deforms the baseline mesh. While this maintains the topology and node count and improves efficiency, it is prone to mesh quality degradation or even negative meshes in large-angle rotations, narrow gaps, and near-wall regions, making it difficult to stably reflect real flow characteristics and hindering consistent comparisons across batches of operating conditions.

[0026] To address this, this application proposes a method for adjusting the computational grid of a compressor. Based on obtaining the baseline geometry and the baseline computational grid, a transformation matrix is ​​introduced to describe the relative geometric changes. A neural network model is constructed and trained, and a two-dimensional grid under the target operating condition is directly output without changing the grid topology and the number of nodes. The grid orthogonality, aspect ratio, and scaling ratio are verified and closed-loop constraints are applied. The two-dimensional grids of each section are stacked along a preset direction to form a three-dimensional grid, which is then merged with the fixed geometric grid at the interface boundary to support the efficient, stable, and comparable computational requirements of multi-stage compressor multi-operating condition simulation.

[0027] Figure 1 A flowchart illustrating the compressor computational grid adjustment method provided in this application. Figure 1 As shown, this embodiment provides an exemplary method for adjusting the computational grid of a compressor, including the following steps: S1. Obtain the reference geometry of the compressor under reference operating conditions and its corresponding reference calculation grid, wherein the reference calculation grid includes a three-dimensional grid formed by stacking multiple two-dimensional grids along a preset direction; S2. Obtain the quality parameters of the benchmark computational grid in step S1, and apply quality constraints to the benchmark computational grid based on the quality parameters and preset conditions. S3. Based on the reference calculation grid that meets the preset conditions in step S2 and the transformation matrix used to adjust the reference geometry to the target working condition, construct and train a neural network model to obtain the mapping of the two-dimensional grid from the reference working condition to the target working condition. S4. Input the mapping in step S3 and the two-dimensional mesh under the benchmark working condition into the trained neural network model to output the two-dimensional mesh under the target working condition, while keeping the topology and the number of mesh nodes unchanged. S5. Based on preset conditions, perform quality verification on the two-dimensional mesh under the target working condition in step S4, and adjust the parameters of the neural network model according to the verification results until the two-dimensional mesh under the target working condition output in step S4 meets the preset conditions. S6. Overlay the two-dimensional mesh under the target working condition output by the neural network model along a preset direction to form a three-dimensional mesh under the target working condition.

[0028] Figure 2 A flowchart illustrating the further compressor computational grid adjustment method provided in this application. Figure 2 As shown, the method also includes step S7, which involves, after forming the three-dimensional mesh under the target operating condition in step S6, merging the three-dimensional mesh of the adjustable geometry in the reference geometry with the three-dimensional mesh of the fixed geometry at a predefined interface boundary to obtain a complete three-dimensional computational mesh for the compressor.

[0029] Figure 3 A schematic diagram of a compressor computational grid adjustment device 10 provided in this application. Figure 3 As shown, the device includes a reference mesh acquisition module 11, used to acquire the reference geometry and its corresponding reference computational mesh under the reference working condition, and to acquire the quality parameters of the reference computational mesh for quality constraints; a mapping model training module 12, used to construct and train a neural network model based on the reference computational mesh that meets preset conditions and the transformation matrix used to adjust the reference geometry to the target working condition, so as to acquire the mapping of the two-dimensional mesh from the reference working condition to the target working condition; a mesh quality constraint unit 13, used to perform quality verification on the two-dimensional mesh under the target working condition, and record the default excess amount and the mesh position corresponding to the maximum default excess amount when the preset conditions are not met; a closed-loop update control unit 14, used to add the quality penalty term to the model training process to form the update loss, update the parameters of the neural network model and re-output the two-dimensional mesh under the target working condition until the preset conditions are met or the number of iterations reaches the preset upper limit; and a target mesh generation module 15, used to output the two-dimensional mesh under the target working condition while keeping the topology and the number of mesh nodes unchanged, and to stack them along a preset direction to form a three-dimensional mesh. The target network generation module 15 is also used to merge the three-dimensional mesh of the adjustable geometry in the reference geometry with the three-dimensional mesh of the fixed geometry at a predefined interface boundary to obtain a complete three-dimensional computational mesh of the compressor.

[0030] In all the exemplary embodiments described above, a multi-stage compressor can be used as an application scenario. The following detailed explanation of the method steps in the exemplary embodiments will focus on a multi-stage compressor as a specific application scenario.

[0031] In multi-stage compressor scenarios, the reference geometry can be understood as the overall machine geometry under reference operating conditions, such as design speed, design mass flow rate, and design inlet guide vane angle. This structure typically includes two main parts: fixed geometry and adjustable geometry. Fixed geometry can include the casing, inner and outer endwalls, impeller, non-adjustable moving blades, and stationary blades, whose shapes remain unchanged under different operating conditions. Adjustable geometry can include adjustable inlet guide vanes (IGVs) or variable stator vanes (VSVs) at each stage. Without altering the geometric topology, IGVs and VSVs are generally allowed to change their attitude or relative position according to operating parameters, such as guide vane installation angle. As a unified reference for neural network training and inference, the reference geometry requires clearly defined geometric elements such as blade profiles, endwall shapes, blade height distribution, and periodic surface positions, while preserving the cross-sectional sequences and connection relationships used for 3D reconstruction.

[0032] Figure 4 A schematic diagram of the adjustable guide vane provided in this application. Figure 4 As shown, this adjustable guide vane serves as the adjustable geometry within the reference geometry. It is adjusted from guide vane 1 under the reference operating condition to guide vane 2 under the target operating condition. The affine transformation involved in this process is a rotational transformation centered on rotation center 3. Its transformation matrix depends on the angle of the adjustable guide vane, i.e.: , Wherein, the transformation matrix T n In this context, A is the affine matrix of the above process, n is the order of the target working conditions, and α is the angle of the guide vane under the specific working condition.

[0033] When meshing the reference geometry of the aforementioned multi-stage compressor under reference operating conditions, a structured mesh is preferred to facilitate boundary layer resolution and periodic surface processing. In the two-dimensional section, a (i, j) regular topology can be used, with nodes arranged along and across the flow direction. A denser wall orthogonal mesh is used in the near-wall region to meet the dimensionless wall distance and layer requirements. The three-dimensional mesh can be obtained by overlaying multiple two-dimensional cross-sectional meshes along a preset direction. When there are multiple rows of blades or multiple stages of channels, mesh blocks can be divided into rows or stages, with periodic surfaces and interface boundaries set at the transition points. For unstructured meshing, triangular, tetrahedral, or polyhedral elements can be used, but the orthogonality, aspect ratio, and scaling ratio of the elements must still meet the quality threshold in the near-wall and gap regions. Regardless of the meshing type used, the reference calculation mesh should maintain consistent node numbers and connectivity across different cross-sections to allow for subsequent adjustments to node coordinates while keeping the topology and number of nodes unchanged.

[0034] In multi-stage compressors, the preset direction for stacking two-dimensional meshes is typically the blade height direction, also known as the spanwise or radial direction, i.e., the direction from the hub / inner endwall to the outer casing / outer endwall. To support two-dimensional to three-dimensional stacking, several characteristic sections need to be selected along the blade height direction. For example, sections can be selected at intervals of 10% of the total blade height. Two-dimensional meshing is performed at each section, and the correspondence with adjacent sections is recorded. During stacking, the same node indexes and element connection order are maintained to ensure the continuity and comparability of the three-dimensional mesh. For multi-stage structures, periodic boundaries or non-conformal interfaces can be set between stages and between dynamic and static components as predefined interfaces. After stacking in the preset direction, the overall interface docking and assembly are performed.

[0035] Regarding boundary and interface settings, the inlet boundary can be defined as a total temperature / total pressure / flow direction angle boundary, and the outlet boundary can be defined as a constant pressure or mass flow rate boundary. For solid walls, a no-slip boundary is used, and the height and number of the first layer are set according to the requirements of the turbulence model. Periodic surfaces are used to reduce the computational domain to a single channel or several channels. In multi-stage assembly, conformal interfaces (one-to-one node correspondence) or non-conformal interfaces (achieving continuity of quantities through interpolation / matching) can be selected as needed, and the geometric position and numbering mapping relationship of the mesh blocks are kept stable at the interface to facilitate the subsequent docking and merging of adjustable and fixed geometric meshes.

[0036] Once the baseline computational grid for the multi-stage compressor under reference operating conditions is determined, the baseline grid acquisition module 11 calculates and acquires the quality parameters of the baseline computational grid, including at least grid orthogonality, grid aspect ratio, and grid scaling ratio. For clarity, the above quality parameters are defined first.

[0037] Mesh orthogonality is used to characterize the degree to which the angle between two sets of parameter lines (i, j directions) of a mesh element "moves away from degradation". For structured quadrilateral elements, the angle between the element vertices is used as the basis for this property. remember It is acceptable , The minimum value at each of the four corners of a cell is taken as the orthogonality of that cell; a larger orthogonality value indicates less degradation. For triangular / polygonal meshes, the minimum value of the included angle between adjacent sides can be taken, or an equivalent orthogonality quality index can be used as a substitute.

[0038] The aspect ratio (AR) of a mesh characterizes the degree of scale difference between the two principal directions of a cell, and can be defined as follows: , The smaller the value, the better. For unstructured meshes, this can be achieved using the ratio of the maximum side length to the minimum height or other equivalent definitions.

[0039] The mesh stretching / expansion ratio characterizes the dimensional smoothness of adjacent elements along the same direction. It is calculated separately along the i and j directions. , The larger of the two values ​​is taken as the scaling ratio for that position, and the smaller the value, the better.

[0040] In the above embodiments, a feasible preset condition includes at least the following three preset thresholds: mesh orthogonality greater than 20°, mesh aspect ratio less than 500, mesh scaling ratio less than 5, and all mesh cell areas are positive, i.e., no negative meshes. When all cells meet the above thresholds, the quality parameters of the baseline computational mesh are determined to meet the preset conditions. It should be noted that during the baseline computational mesh division and any subsequent smoothing redistribution or densification adjustment process, area positivity checks and local corrections are set. For cells with negative or near-zero areas, local node redistribution or boundary layer parameter rollback is performed to ensure no negative meshes. If engineering constraints (such as interface alignment, periodic requirements) make it difficult to meet the unified threshold locally, an acceptable quality baseline can be determined through the "optimal value" path. For example, when the values ​​of various quality parameters do not meet the above preset thresholds, the quality parameters are iterated by adjusting the mesh division strategy. Optional mesh division strategies include smoothing redistribution of mesh nodes, local mesh densification, boundary layer parameter correction, etc., so that the quality parameters of the re-divided baseline computational mesh are close to the above preset thresholds by adjusting the division strategy. When using a mesh generation strategy to adjust the baseline geometry under the baseline working condition, the iteration stops if the number of mesh nodes exceeds a preset limit, and the current quality parameters are taken as the optimal values. The preset limit is determined by the computational scale and solver resources; the optimal value refers to the best achievable value of the combination of quality indicators obtained through a finite number of iterations under the constraint that the number of nodes does not exceed the limit.

[0041] To facilitate subsequent geometric transformations and mesh adjustments, during the stage of obtaining the baseline computational mesh through baseline geometric mesh generation, the cross-section parameterization and computational domain definition are pre-completed. This includes the two-dimensional airfoil geometric parameters of each cross-section (chord length, thickness distribution, installation angle, stacking lines, etc.), the computational domain range (inlet / outlet cross-section position, channel height, channel shape), and a transformation matrix used to describe the relative changes from the baseline geometry to the target operating condition (e.g., rotation around the pivot, necessary translation / scaling, and computational domain fine-tuning). This transformation matrix can be directly generated from mechanism parameters (such as guide vane angle, pivot position, domain scale / inclination) or obtained based on least-squares fitting of the baseline / target boundary point set. The above information will serve as the input conditions for the neural network model or the basis for the initial analytical transformation, enabling stable and controllable adjustment of the coordinates of the two-dimensional mesh nodes of each cross-section while maintaining the topology and number of nodes, and finally stacking them along the blade height direction to form the three-dimensional mesh of the target operating condition.

[0042] In the embodiments of this application, the neural network model in step S3 above can use a fully connected neural network (FCNN) model. This model is used to learn and output the two-dimensional grid node coordinates or node displacements under the target working condition without changing the grid topology and the number of grid nodes, corresponding one-to-one with the grid nodes under the reference working condition. Specifically, the input layer of this neural network model includes geometric and grid information corresponding to different cross-sectional heights, as well as parameters for describing the changes in working condition. These parameters include: a two-dimensional airfoil, a set of parameters characterizing the geometry of the airfoil at that cross-section, which can be profile sampling points, spline control points, or other geometric features; a two-dimensional computational domain, a set of parameters describing the flow channel boundary and the scale / shape of the computational domain at that cross-section; and the reference two-dimensional grid node coordinates, i.e., the grid nodes (x, y, x) under the reference working condition at that cross-section. i ,y i To maintain consistent node numbering and connectivity, a transformation matrix is ​​used to adjust the baseline geometry to the target working condition. This matrix can be a combination of affine transformations (rotation, scaling, translation, shearing) and is input after being flattened by cross-section. The above input layers can be fed in point-by-point at the node level or in batches at the cross-section level, without affecting the substance of this application.

[0043] Meanwhile, the hidden layers of FCNN consist of several fully connected layers, and the output layer outputs the two-dimensional coordinates (x, y, y) of the target working condition at each input node. i ',y i '), or output nodal displacement (Δx) i , Δy i ) and through (x i ',y i ') = (x i ,y i ) + (Δx i ,Δy i The target coordinates are obtained. Regardless of the output format, FCNN ensures a one-to-one correspondence between the node numbers of the grid under the target operating condition and the reference calculation grid, thus maintaining the topology and the number of nodes unchanged. For multi-stage compressors, several sections are selected along the blade height direction, and inference can be repeated on each section using shared weights, or sub-models / sub-heads can be configured for key sections. For nodes located on solid walls / geometric boundaries, boundary projection can be performed after output to fit the boundary curves / segments under the target operating condition, thus ensuring geometric consistency, and this operation does not change the total number of nodes or connectivity.

[0044] In the hidden layers of the aforementioned FCNN, one optional implementation uses ReLU (Rectified Linear Unit) as the activation function to improve nonlinear representation capabilities and ensure training stability and convergence efficiency. During model training, mean squared error is selected as the basic loss function to measure the deviation between the model output and the target value, as detailed below: When the output layer directly provides the target node coordinates, the basic loss is the average of the squared differences between the model-predicted grid node coordinates and the grid node coordinates under the target working condition, calculated component by component. When the output layer provides the node displacement, the basic loss is the average of the squared differences between the model-predicted node displacement and the node displacement under the target working condition, calculated component by component. These two supervised objectives are equivalent, both satisfying the condition that the input to the mean squared error function is the model-predicted grid node coordinates and the grid node coordinates (or corresponding displacements) under the target working condition, and the output is the squared mean of the differences between the inputs.

[0045] The training process of FCNN is described below. The training data consists of sample pairs between a baseline computational grid and the computational grid under the target working condition. Samples can be batched by node or by cross-section. The base loss is the average of all nodes in the current batch. For example, a fully connected neural network is used to train 100 sets of samples, with 85% used as the training set and 15% as the validation set. The mean squared error (MSE) is used as the base loss to measure the predicted grid node coordinates (or node displacements) and the target grid node coordinates (or node displacements). When the MSE of the validation set reaches a preset target value and / or improves less than a preset threshold within several consecutive rounds, training convergence is determined, and the model parameters with the lowest validation error are saved as the final model. In other embodiments, without changing the definition of the base loss, a grid quality-related penalty term can be added, or a weighted MSE can be used to emphasize the importance of near-wall / narrow gap regions. It should be understood that such expansion and closed-loop updates are not intended to limit the independent claims of this application. In the above process, the true values ​​of the computational grid samples under the target working condition can be obtained through a high-quality re-mesh or a traditional deformation process verified by engineering, and used as a supervision signal.

[0046] Figure 5 The flowchart for quality verification and neural network parameter adjustment provided for this application is shown below. Figure 5 As shown, this process is used for a detailed explanation of step S5 above, and it mainly includes the following steps: S51. Obtain the quality parameters of the two-dimensional mesh under the target working condition. The quality parameters include at least the mesh orthogonality, mesh aspect ratio, and mesh scaling ratio. S52. Compare the quality parameters of the two-dimensional mesh under the target working condition with the preset conditions; S53. Determine whether all values ​​of the quality parameters of the two-dimensional mesh under the target working condition meet the preset conditions. If yes, determine that the verification result is passed and output the current two-dimensional mesh. If no, proceed to step S54. S54. When the quality parameter of any two-dimensional grid does not meet the preset condition, record the excess amount of the quality parameter that does not meet the preset condition relative to the preset condition, and the grid position corresponding to the maximum excess amount of the default. S55. Construct a quality penalty term based on the recorded information, and add the quality penalty term to the training process of the neural network model to iteratively update the two-dimensional mesh under the target working condition, and re-verify the quality, and execute step S52. S56. When the number of iterations in step S55 reaches the preset upper limit, output the current two-dimensional mesh.

[0047] In steps S51 to S56 above, quality parameters are calculated on each two-dimensional mesh cell and compared with preset conditions item by item. Preferred preset conditions include orthogonality ≥20°, aspect ratio ≤500, scaling ratio ≤5, and further, positive cell area. When engineering constraints cause a certain location on the reference mesh to fail to meet a unified threshold, the measured value at that location is used as the constraint baseline. For cells that fail to meet the standards, the excess amount of each index and the location of the cell with the largest excess amount are recorded, which can be identified by cell index, mesh (i, j) number, or cell centroid coordinates. The quality penalty term can be weighted with the basic loss, such as mean square error, to form an update loss. Based on this, small-step parameter updates are performed on the neural network model, and the two-dimensional mesh is re-output. The quality comparison is then repeated until all cells meet the preset conditions or the number of iterations reaches the preset upper limit. For nodes located on solid walls / interfaces, it is preferable to perform boundary projection after each update to maintain geometric attachment without changing the mesh topology and the number of nodes.

[0048] Figure 6 This is a schematic diagram of the reference calculation grid for the adjustable guide vane under the reference operating condition provided in this application. Figure 6 As shown in the figure, the adjustable guide vane has been meshed under the reference working condition, and the mesh quality meets the expected requirements. The figure shows the two-dimensional mesh of the guide vane at a certain height h, and it can be seen intuitively that there are no negative mesh elements in the two-dimensional mesh.

[0049] Figure 7 This is a schematic diagram of the neural network output provided in this application. Figure 7 As shown in the figure, the adjustable guide vane is the one described above. Figure 6 The target operating condition achieved after adjusting the angle of the adjustable guide vane. After the neural network training process provided in this embodiment, the neural network outputs... Figure 7 The two-dimensional mesh under the target working condition, compared to the one under the reference working condition. Figure 6The results show that the number of grid nodes remains unchanged, only the node coordinates change, and there are no negative grids.

[0050] In summary, this specific embodiment takes a multi-stage compressor as a classic application scenario and proposes a method and apparatus for adjusting the compressor computational grid: A baseline geometry and a baseline computational grid are obtained under a baseline operating condition; grid quality parameters are extracted and preset conditions are set; based on a transformation matrix describing the change from the baseline geometry to the target operating condition, a fully connected neural network is constructed and trained; the two-dimensional airfoil, computational domain range, baseline grid nodes, and transformation matrix are input; and a target two-dimensional grid corresponding one-to-one with the baseline nodes is output; quality verification and closed-loop updates are performed on the output grid until the preset conditions are met; the two-dimensional grids of each section are superimposed along the blade height direction to form a three-dimensional grid, and then merged with the fixed geometric grid at the interface boundary to obtain the overall machine computational grid.

[0051] The technical solution of this application can directly generate the target working condition mesh without changing the mesh topology and the number of nodes, avoiding re-meshing for each working condition and improving preprocessing efficiency; quality constraints and closed-loop updates control orthogonality, aspect ratio and scaling ratio during the generation stage, reducing the risk of negative mesh and non-convergence, and improving computational stability; maintaining the consistency of node numbering and interface relationship facilitates cross-working condition comparison and boundary condition reuse; the process is clear, easy to integrate with existing simulation systems and supports batch working conditions.

[0052] The above embodiments are provided only to help understand the methods, structures, and core ideas of this application. Those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the scope of protection of the claims.

Claims

1. A method for adjusting the computational grid of a compressor, characterized in that, include: Obtain the reference geometry of the compressor under reference operating conditions and its corresponding reference calculation grid, wherein the reference calculation grid is formed by stacking multiple two-dimensional grids under reference operating conditions along a preset direction; Obtain the quality parameters of the benchmark computing grid, and apply quality constraints to the benchmark computing grid based on the quality parameters and preset conditions; Based on the reference computing grid that meets the preset conditions and the transformation matrix used to adjust the reference geometry to the target working condition, a neural network model is constructed and trained to obtain the mapping of the two-dimensional grid from the reference working condition to the target working condition; The mapping and the two-dimensional mesh under the benchmark condition are input into the trained neural network model to output the two-dimensional mesh under the target condition, while keeping the topology and the number of mesh nodes unchanged. The quality of the two-dimensional mesh under the target working condition is verified based on the preset conditions, and the parameters of the neural network model are adjusted according to the verification results until the two-dimensional mesh under the target working condition meets the preset conditions. The two-dimensional mesh under the target working condition output by the neural network model is superimposed along the preset direction to form a three-dimensional mesh under the target working condition.

2. The compressor computational grid adjustment method as described in claim 1, characterized in that, The compressor is a multi-stage compressor, and its reference geometry includes at least one adjustable guide vane. The reference calculation grid is generated by structured partitioning to ensure no negative grid. The preset direction is the blade height direction. The target operating condition is achieved by adjusting the angle of the at least one adjustable guide vane.

3. The compressor computational grid adjustment method as described in claim 1, characterized in that, The quality parameters of the benchmark computing grid include at least grid orthogonality, grid aspect ratio, and grid scaling ratio. The preset condition is that all values ​​of the quality parameters meet preset thresholds, or that all values ​​of the quality parameters are optimal values ​​when the number of grid nodes does not exceed a preset limit. Specifically, when the values ​​of the quality parameters do not meet the preset threshold, the mesh partitioning strategy is adjusted to iterate the quality parameters so that the values ​​of the quality parameters are close to the preset threshold. When the number of mesh nodes exceeds the preset limit, the iteration stops, and the current values ​​of the quality parameters are taken as the optimal values.

4. The compressor computational grid adjustment method as described in claim 1, characterized in that, The neural network model is a fully connected neural network model; The input layer of the fully connected neural network includes two-dimensional leaf profiles with different cross-sectional heights, a two-dimensional computational domain range under the reference working condition, two-dimensional grid node coordinates under the reference working condition, and the transformation matrix. The output layer outputs the two-dimensional grid node coordinates or grid node displacements under the target working condition corresponding to different cross sections, and they correspond one-to-one with the grid nodes under the reference working condition.

5. The compressor computational grid adjustment method as described in claim 4, characterized in that, The hidden layers of the fully connected neural network use ReLU as the activation function; The basic loss during model training is the mean squared error function. The input of the mean squared error function is the grid node coordinates predicted by the model and the grid node coordinates under the target working condition, and the output is the squared mean of the difference between the input values.

6. The compressor computational grid adjustment method as described in claim 1, characterized in that, The quality verification and the adjustment of the parameters of the neural network model based on the verification results include: Obtain the quality parameters of the two-dimensional mesh under the target working condition. The quality parameters include at least mesh orthogonality, mesh aspect ratio, and mesh scaling ratio. The quality parameters of the two-dimensional mesh under the target working condition are compared with the preset conditions; When all the quality parameters of the two-dimensional mesh under the target working condition meet the preset conditions, the verification result is determined to be passed, and the current two-dimensional mesh is output. When the quality parameter of any two-dimensional grid does not meet the preset condition, record the excess amount of the quality parameter that does not meet the preset condition relative to the preset condition and the grid position corresponding to the maximum excess amount of the default. A quality penalty term is constructed based on the recorded information, and the quality penalty term is added to the training process of the neural network model to iteratively update the two-dimensional mesh under the target working condition, and the quality is re-verified; Repeat the above steps until the quality parameters of all two-dimensional meshes meet the preset conditions or the number of iterations reaches the preset upper limit, then output the current two-dimensional mesh.

7. The compressor computational grid adjustment method as described in claim 1, characterized in that, Also includes: When outputting the two-dimensional mesh under the target working condition, perform boundary projection on the mesh nodes located on the solid wall boundary to fit the geometric boundary under the target working condition; After forming the three-dimensional mesh under the target operating condition, the three-dimensional mesh of the adjustable geometry in the reference geometry is docked and merged with the three-dimensional mesh of the fixed geometry at a predefined interface boundary to obtain a complete three-dimensional computational mesh of the compressor.

8. A compressor computational grid adjustment device, characterized in that, include: The reference mesh acquisition module is used to acquire the reference geometry and its corresponding reference calculation mesh under the reference working condition, and to acquire the quality parameters of the reference calculation mesh for quality constraints. The mapping model training module is used to construct and train a neural network model based on the reference computing grid that meets preset conditions and the transformation matrix used to adjust the reference geometry to the target working condition, so as to obtain the mapping of the two-dimensional grid from the reference working condition to the target working condition. Mesh quality constraint element is used to perform quality verification on the two-dimensional mesh under the target working condition, and to record the amount of violation and excess and the mesh position corresponding to the maximum amount of violation when the preset conditions are not met. The closed-loop update control unit is used to add a quality penalty term to the model training process to form an update loss, update the parameters of the neural network model and re-output the two-dimensional mesh under the target working condition until the preset conditions are met or the number of iterations reaches the preset upper limit. The target mesh generation module is used to output a two-dimensional mesh under the target working condition while keeping the topology and the number of mesh nodes unchanged, and to stack them along a preset direction to form a three-dimensional mesh.

9. The compressor computational grid adjustment device as described in claim 8, characterized in that, The neural network model is a fully connected neural network; The input layer of the fully connected neural network includes two-dimensional leaf profiles with different cross-sectional heights, two-dimensional computational domain range under the reference working condition, two-dimensional grid node coordinates under the reference working condition, and the transformation matrix. The output layer outputs the two-dimensional grid node coordinates or grid node displacements under the target working condition corresponding to different cross sections, and corresponds one-to-one with the grid nodes under the reference working condition. The hidden layers of the fully connected neural network use ReLU as the activation function. The basic loss during model training is the mean squared error function, which is weighted and summed with the quality penalty term provided by the grid quality constraint unit to form the update loss.

10. A computer-readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the compressor computational grid adjustment method according to any one of claims 1 to 7.

Images