A method and related apparatus for generating four-dimensional spatiotemporal adaptive grids for meteorological numerical forecasting
By using a four-dimensional spatiotemporal adaptive grid generation method, the resolution of the meteorological numerical forecast grid is dynamically adjusted, which solves the matching problem between the static heterogeneity of the underlying surface and the dynamic changes in weather conditions, thereby improving forecast accuracy and resource utilization efficiency.
Patent Information
- Application Number
- CN202611240657.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-17
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-08-17
AI Technical Summary
Existing meteorological numerical forecasting methods are unable to comprehensively reflect the static heterogeneity of the underlying surface and the dynamic changes in weather conditions, resulting in inaccurate grid resolution configuration in areas where complex underlying surfaces and severe weather processes overlap in time and space.
A four-dimensional spatiotemporal adaptive grid generation method is adopted. By obtaining static and dynamic resolution weight fields, a comprehensive resolution control field is formed using a nonlinear fusion function to dynamically adjust the computational grid resolution. Combined with a hierarchical grid data structure and a conservation remapping operator, the grid can be refined or coarsened.
It improves the matching degree between the configuration of meteorological numerical forecast grid and actual computing needs, enhances local computing accuracy and resource utilization efficiency, and reduces unnecessary high-resolution regional computing load.
Smart Images

Figure CN122797170B_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of meteorological numerical forecasting technology, specifically relating to a method and related apparatus for generating a four-dimensional spatiotemporal adaptive grid for meteorological numerical forecasting. Background Technology
[0002] Numerical weather prediction typically involves numerically integrating atmospheric motion equations on a discretized spatial grid. The spatial resolution of the computational grid directly impacts the accuracy, numerical stability, and computational resource consumption of weather forecasts. Existing meteorological numerical prediction models often employ fixed-resolution grids, static nested grids, or pre-defined variable-resolution grids. To improve forecast accuracy in local areas, it is usually necessary to increase the grid resolution over a larger area, leading to a corresponding increase in computational load. Some adaptive grid methods refine or coarsen the computational grid based on local numerical errors, flow field gradients, topographic conditions, or weather conditions, thus concentrating computational resources on areas requiring higher resolution.
[0003] However, existing methods typically adjust the grid based on a single type of factor in the underlying surface characteristics or weather conditions, or process static and dynamic factors separately and then simply overlay them. This makes it difficult to reflect the mutual reinforcement effect when complex underlying surfaces and severe weather processes overlap in time and space, resulting in a mismatch between the grid resolution configuration of the corresponding region and the actual computational needs. Summary of the Invention
[0004] This application addresses the technical problem that existing meteorological numerical forecasting grids are unable to effectively reflect the static heterogeneity of the underlying surface and the dynamic changes in weather conditions, leading to inaccurate grid resolution configuration in areas where complex underlying surfaces and severe weather processes overlap in time and space. It provides a method and related apparatus for generating a four-dimensional spatiotemporal adaptive grid for meteorological numerical forecasting.
[0005] To achieve the above objectives, this application adopts the following technical solution: Firstly, this application proposes a four-dimensional spatiotemporal adaptive grid generation method for meteorological numerical forecasting, including: Obtain a static resolution weight field determined based on the underlying surface data of the forecast area, wherein the static resolution weight field is used to characterize the static heterogeneity of the underlying surface of the forecast area; During the integration process of the meteorological numerical forecasting model, the meteorological state field of the forecast area at the current integration time is obtained, and the dynamic resolution weight field is determined based on the meteorological state field. The static resolution weight field and the dynamic resolution weight field are fused using a nonlinear fusion function to obtain a comprehensive resolution control field. The nonlinear fusion function satisfies the following: when one term of the static resolution weight field or the dynamic resolution weight field increases while the other term remains unchanged, the comprehensive resolution control field remains unchanged or increases; when both the static resolution weight field and the dynamic resolution weight field increase together, the increment of the comprehensive resolution control field is greater than the sum of the increments produced by the individual increases of the static resolution weight field and the dynamic resolution weight field. The target grid resolution field is determined based on the comprehensive resolution control field. The computational grid of the prediction region is refined or coarsened according to the target grid resolution field to obtain a four-dimensional spatiotemporal adaptive grid that is updated with the integration time.
[0006] Furthermore, the method for determining the static resolution weight field includes: Obtain digital elevation model data and land use and land cover type data for the forecast area; The terrain complexity weight is determined based on the digital elevation model data, and the land use mixing weight is determined based on the land use and land cover type data. The terrain complexity weights and land use hybridity weights are aggregated using a monotonic aggregation operator to obtain the static resolution weight field.
[0007] Furthermore, the aggregation operator that maintains monotonicity satisfies the following condition: when either the terrain complexity weight or the land use hybridity weight increases, the static resolution weight field remains unchanged or increases.
[0008] Further, determining the dynamic resolution weight field based on the meteorological state field includes: Based on the meteorological state field, at least one meteorological diagnostic quantity related to high-impact weather processes is determined; The meteorological diagnostic data are normalized. The normalized meteorological diagnostic data are aggregated using a monotonic aggregation operator to obtain the dynamic resolution weight field.
[0009] Furthermore, the nonlinear fusion function is an addition-multiplication fusion function, a power-law cooperative fusion function, a product-amplification fusion function, an exponential cooperative fusion function, or a learnable fusion model.
[0010] Furthermore, the nonlinear fusion function is a learnable fusion model; The learnable fusion model is trained under supervision based on historical weather forecast errors, and monotonicity constraints and collaborative enhancement constraints are applied during the training process to ensure that the learnable fusion model satisfies the conditions of the nonlinear fusion function.
[0011] Furthermore, the target grid resolution field is composed of the target grid cell scale corresponding to each spatial location in the prediction region; Determining the target grid resolution field based on the comprehensive resolution control field includes: The integrated resolution control field is converted into the target grid cell scale corresponding to each spatial location in the forecast area by a monotonically non-increasing mapping, so that when the integrated resolution control field increases, the corresponding target grid cell scale remains unchanged or decreases.
[0012] Furthermore, the computational grid is formed based on a pre-built and stored hierarchical grid data structure, which includes multiple grid resolution levels; The computational grid of the prediction region is refined or coarsened according to the target grid resolution field, including: The target grid level corresponding to each spatial location is determined based on the target grid resolution field, and the computational grid is refined or coarsened by activating or suspending the computational grid cells corresponding to the grid resolution level.
[0013] Furthermore, when refining or coarsening the computational grid by activating or suspending the computational grid cells corresponding to the grid resolution level, the process is executed according to multiple grid update trigger levels. The multiple grid update trigger levels correspond to the characteristic time scales of weather processes of different scales, and each has a corresponding grid update time interval and control field change threshold. When the grid update time corresponding to any grid update trigger level arrives: Determine the amount of change in the integrated resolution control field relative to the previous grid update time; For computational grid cells whose change exceeds the corresponding control field change threshold, the corresponding target grid level is determined based on the target grid resolution field, and the computational grid cells at the corresponding grid resolution level are activated or dormant.
[0014] Furthermore, after refining or coarsening the computational grid of the forecast area according to the target grid resolution field, the method further includes: using a conservation remapping operator that satisfies variable integral conservation to remap the meteorological state variables constituting the meteorological state field on the computational grid before refinement or coarsening to the computational grid after refinement or coarsening.
[0015] Secondly, this application proposes a four-dimensional spatiotemporal adaptive grid generation system for meteorological numerical forecasting, comprising: The static resolution weight field module is used to obtain a static resolution weight field determined based on the underlying surface data of the forecast area. The static resolution weight field is used to characterize the static heterogeneity of the underlying surface of the forecast area. The dynamic resolution weight field module is used to obtain the meteorological state field of the forecast area at the current integration time during the integration process of the meteorological numerical forecast model, and to determine the dynamic resolution weight field based on the meteorological state field. The fusion module is used to fuse the static resolution weight field and the dynamic resolution weight field through a nonlinear fusion function to obtain a comprehensive resolution control field; wherein, the nonlinear fusion function satisfies the following: when either the static resolution weight field or the dynamic resolution weight field increases while the other remains unchanged, the comprehensive resolution control field remains unchanged or increases; when both the static resolution weight field and the dynamic resolution weight field increase together, the increment of the comprehensive resolution control field is greater than the sum of the increments produced by the individual increases of the static resolution weight field and the dynamic resolution weight field. The target grid resolution field module is used to determine the target grid resolution field based on the comprehensive resolution control field. The update module is used to refine or coarse the computational grid of the prediction region according to the target grid resolution field, so as to obtain a four-dimensional spatiotemporal adaptive grid that updates with the integration time.
[0016] Thirdly, this application proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-mentioned four-dimensional spatiotemporal adaptive grid generation method for meteorological numerical forecasting.
[0017] Fourthly, this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described four-dimensional spatiotemporal adaptive grid generation method for meteorological numerical forecasting.
[0018] Compared with the prior art, this application has the following beneficial effects: This application proposes a four-dimensional spatiotemporal adaptive grid generation method for meteorological numerical weather prediction. It obtains a static resolution weight field characterizing the static heterogeneity of the underlying surface in the forecast area, and determines a dynamic resolution weight field based on the meteorological state field at the current integration time during the integration process of the meteorological numerical weather prediction model. This method can reflect the spatial differences of the underlying surface and the changes in weather conditions over time. By fusing the static and dynamic resolution weight fields using a nonlinear fusion function that satisfies monotonicity and synergistic enhancement conditions, the overall resolution control field does not decrease when either weight field increases. Furthermore, it achieves a higher enhancement control effect than either weight field acting separately when complex underlying surfaces and severe weather processes overlap spatiotemporally and temporally, thereby enabling more accurate identification of areas with high grid resolution requirements. Further, by determining the target grid resolution field based on the comprehensive resolution control field, and refining or coarsening the computational grid accordingly, the computational grid can be dynamically updated as the weather conditions evolve. This improves the local grid resolution in areas with complex underlying surfaces and overlapping severe weather processes, while avoiding maintaining unnecessary high resolution in areas with low resolution requirements. This improves the matching degree between the meteorological numerical forecast grid configuration and actual computational needs, and is conducive to balancing the local computational accuracy of numerical forecasts with the efficiency of computational resource utilization.
[0019] This application also proposes a four-dimensional spatiotemporal adaptive grid generation system for meteorological numerical forecasting, an electronic device, and a computer-readable storage medium, which possess the advantages of the aforementioned four-dimensional spatiotemporal adaptive grid generation method for meteorological numerical forecasting. Attached Figure Description
[0020] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0021] Figure 1 This is a schematic diagram of the four-dimensional spatiotemporal adaptive grid generation method for meteorological numerical forecasting proposed in this application. Figure 2 This is another schematic diagram of the four-dimensional spatiotemporal adaptive grid generation method for meteorological numerical forecasting in this application; Figure 3 This is a schematic diagram of a four-dimensional spatiotemporal adaptive grid generation system for meteorological numerical forecasting in this application. Detailed Implementation
[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0023] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.
[0024] This application proposes a four-dimensional spatiotemporal adaptive grid generation method and related apparatus for meteorological numerical weather prediction. During the integration process of a meteorological numerical weather prediction model, the spatial resolution of the computational grid is collaboratively controlled based on the static heterogeneity of the underlying surface and the dynamic changes in weather conditions within the forecast area. The computational grid generated by this method not only changes with spatial location but also continuously updates over integration time, thereby prioritizing the allocation of limited computational resources to areas with more complex underlying surfaces, stronger weather processes, or where these two aspects overlap spatiotemporally.
[0025] like Figure 1 The diagram shown is a schematic representation of a four-dimensional spatiotemporal adaptive grid generation method for meteorological numerical weather prediction proposed in this application, which may include: S101, Obtain the static resolution weight field determined based on the underlying surface data of the forecast area.
[0026] The static resolution weight field corresponds to the spatial location within the forecast region and is used to characterize the static heterogeneity of the underlying surface at each spatial location. The more complex the underlying surface and the more significant the spatial differences, the higher the corresponding static resolution weight can be; the more homogeneous the underlying surface, the lower the corresponding static resolution weight can be. The static resolution weight field can remain valid during a single forecast integration process or can be regenerated when the underlying surface data is updated.
[0027] S102, during the integration process of the meteorological numerical forecasting model, the meteorological state field of the forecast area at the current integration time is obtained, and the dynamic resolution weight field is determined based on the meteorological state field.
[0028] The meteorological state field comprises one or more meteorological state variables generated by the meteorological numerical weather prediction model on the current computational grid, reflecting the current atmospheric state. Based on the intensity, spatial variation, or evolution of weather processes reflected in the meteorological state field, a dynamic resolution weight field corresponding to the current integration time is determined. This dynamic resolution weight field changes with the meteorological state field and can generate different spatial distributions at different integration times, enabling the computational grid to adjust to the movement, development, and dissipation of weather systems.
[0029] S103, the static resolution weight field and the dynamic resolution weight field are fused by a nonlinear fusion function to obtain a comprehensive resolution control field.
[0030] It should be noted that the nonlinear fusion function maintains a monotonically non-decreasing relationship with respect to the two inputs: when the static resolution weight field increases while the dynamic resolution weight field remains unchanged, the comprehensive resolution control field remains unchanged or increases; when the dynamic resolution weight field increases while the static resolution weight field remains unchanged, the comprehensive resolution control field also remains unchanged or increases.
[0031] This nonlinear fusion function also exhibits synergistic enhancement characteristics. When both the static resolution weight field and the dynamic resolution weight field increase together, the increment generated by the integrated resolution control field is greater than the sum of the increments generated by increasing only the static resolution weight field and increasing only the dynamic resolution weight field. Therefore, when complex underlying surfaces and severe weather processes overlap spatiotemporally, the integrated resolution control field can generate additional synergistic enhancement, giving the overlapping region a higher resolution configuration priority than when the two driving factors act independently, thus preventing the static underlying surface factors and dynamic weather factors from weakening each other after simple averaging.
[0032] S104, Determine the target grid resolution field based on the comprehensive resolution control field.
[0033] The target grid resolution field is used to determine the desired grid resolution for each spatial location in the forecast area at the current integration time. Locations with higher integrated resolution control fields correspond to higher computational resolutions, while locations with lower integrated resolution control fields correspond to lower computational resolutions. This transforms the integrated control requirements arising from static heterogeneity and dynamic weather conditions into actionable grid scale requirements.
[0034] S105, refine or coarse the computational grid of the prediction region according to the target grid resolution field to obtain a four-dimensional spatiotemporal adaptive grid that updates with the integration time.
[0035] Based on the target grid resolution field, the existing computational grid in the forecast area is compared with the target grid resolution. For spatial locations where the current computational resolution is lower than the target resolution, the corresponding computational grid is refined; for spatial locations where the current computational resolution is higher than the target resolution and meets the coarsening conditions, the corresponding computational grid is coarsened; for spatial locations where the current computational resolution matches the target resolution, the corresponding computational grid remains unchanged. After refinement or coarsening, the computational grid corresponding to the current integration time is formed, and the meteorological numerical weather prediction model continues integration on the updated computational grid.
[0036] As integration progresses, the following processes are repeated: acquiring the meteorological state field at the current integration time, updating the dynamic resolution weight field, forming the comprehensive resolution control field, determining the target grid resolution field, and refining or coarsening the computational grid, allowing the computational grid to continuously evolve within the three-dimensional forecast space over time. This enables the system to reflect the static heterogeneity of the underlying surface over the long term, track the dynamic changes of weather processes in real time, and form a synergistically enhanced resolution configuration in areas where complex underlying surfaces overlap with severe weather processes, thus improving the problem that existing grids cannot accurately cover key spatiotemporal regions.
[0037] The present application will be further described in detail below through a more detailed embodiment. For example... Figure 2 The diagram shown is another schematic representation of the principle of the four-dimensional spatiotemporal adaptive grid generation method for meteorological numerical forecasting proposed in this application.
[0038] In one embodiment of this application, the four-dimensional spatiotemporal adaptive grid generation method for meteorological numerical forecasting is applied to the integration process of a meteorological numerical forecasting model. The meteorological numerical forecasting model performs numerical integration of the atmospheric motion equations on a discretized three-dimensional spatial grid, and dynamically adjusts the computational grid resolution for different spatial locations based on the static heterogeneity of the underlying surface and the current weather conditions during the integration time progression, thereby forming a four-dimensional spatiotemporal adaptive grid that varies with both spatial location and integration time. The following formulas use… , The horizontal and vertical coordinates representing spatial locations are... This represents the integration time, and this representation does not impose any restrictions on the actual spatial dimensions or grid coordinate format.
[0039] Existing numerical weather prediction models mostly employ fixed grids with approximately uniform spatial resolution, resulting in similar computational resource consumption across different regions throughout the forecast integration period. For complex terrain areas, areas with mixed land use types, and regions experiencing localized severe weather, uniformly refining the grid across the entire region to improve the computational resolution of key areas leads to a waste of computational resources in flat and uniform areas. Pre-set static nested grids maintain their position and extent unchanged during integration, making it difficult to follow the movement of weather systems. Adaptive grid refinement methods that use local numerical errors or general flow field gradients as triggers struggle to simultaneously reflect the static heterogeneity of the underlying surface, the dynamic changes in meteorological operational diagnostic data, and the synergistic effect of both in the same spatiotemporal region. Therefore, this embodiment sequentially forms a static resolution weight field and a dynamic resolution weight field, which are then fused using a nonlinear fusion function with monotonicity and synergistic enhancement characteristics to obtain a comprehensive resolution control field. Finally, a hierarchical grid data structure, a graded triggering mechanism, and a conserved remapping process are used to update the computational grid.
[0040] In practical applications, this embodiment can be integrated into meteorological numerical forecasting models that employ unstructured or semi-structured grids, and can also be applied to dynamic frameworks employing the finite volume method, finite difference method, or spectral element method, depending on actual needs. The computational grid can be triangular, quadrilateral, hexagonal, or cubic sphere grids. For different meteorological numerical forecasting model frameworks, it is only necessary to adapt the grid data structure and conserved remapping interface in the meteorological numerical forecasting model to the grid update process of this embodiment; there is no need to change the processing relationships between the static resolution weight field, dynamic resolution weight field, integrated resolution control field, and target grid resolution field.
[0041] Before the meteorological numerical weather prediction model is launched, a hierarchical grid data structure with multiple grid resolution levels is pre-constructed and stored, and this hierarchical grid data structure can reside in memory. The hierarchical grid data structure is decoupled from the resolution control field calculated in real time during the integration process of the meteorological numerical weather prediction model, so that the grid update process can be completed through target level lookup and activation or dormancy of corresponding grid cells, without having to re-execute global grid subdivision for each update. Optionally, the hierarchical grid data structure can adopt a quadtree, octree, KD tree, hierarchical grid tree with arbitrary refinement ratio, structured multi-layer grid, or unstructured multi-layer grid, and the grid refinement ratio is not limited to 2.
[0042] As an example, level 0 corresponds to a resolution of 9 km, with a refinement ratio of 2 between adjacent levels. Level 4 corresponds to a resolution of approximately 0.6 km. The total number of levels is... Based on forecast accuracy and computational resource allocation, the runtime system can directly locate the target grid level at the corresponding spatial location based on the target grid side length, and activate or suspend the computational grid cells at the corresponding level. A single level lookup and state switch can achieve constant or near-constant time complexity, for example... .
[0043] In the static driving layer, during the initialization of the meteorological numerical forecasting model, digital elevation model data, land use and land cover type data of the forecast area are acquired, and a static resolution weight field for the entire region is generated based on the aforementioned underlying surface data. The static resolution weight field is formed by aggregating one or more sub-weights that characterize the static heterogeneity of the underlying surface. In one embodiment of this application, the static resolution weight field includes terrain complexity weights. and land use mixing weight .
[0044] The terrain complexity weight remains constant or increases with increasing terrain ruggedness, undulation, and geometric variation. Terrain factors used to calculate the terrain complexity weight may include one or more of the following: terrain curvature, slope variance, Terrain Ruggedness Index (TRI), Topographic Position Index (TPI), slope, aspect variability, local elevation standard deviation, and surface fractal dimension. Specifically, a circle with the center of the computational grid cell and a radius of... Within the analysis window, based on the terrain curvature and slope variance The terrain complexity weight is calculated using the following formula:
[0045] in, Indicates normalization to interval, The weighting coefficients for the corresponding items (as an example, take...) (In practice, adjustments can be made based on terrain features, or a non-additive combination can be used). One method for calculating terrain curvature is the root mean square of the second-order partial derivative of the DEM (Digital Elevation Model) elevation field:
[0046] in, This represents the ground elevation value. Additionally, the shape (e.g., circular, square, or other) and size of the analysis window are also considered. Both the normalization method and the configuration method can be configured and do not constitute a limitation.
[0047] The land use mixing weight remains constant or increases with the increase of the diversity, spatial fragmentation, or density of type boundaries of land use and land cover types. When the underlying surface in the analysis window is a single homogeneous type, the land use mixing weight takes a lower value to allow for coarsening of the computational grid in the corresponding area; when the analysis window contains more land cover types, has a high degree of type mosaic, or has denser type boundaries, the land use mixing weight increases accordingly.
[0048] Specifically, this can be done within the analysis window based on Shannon information entropy. and type boundary density Calculate the land use mixing weights. Where:
[0049] The land use mixing weight is calculated according to the following formula. :
[0050] LULC stands for Land Use and Land Cover. For the first in the window The area proportion of land features These are the corresponding weight coefficients (as an example, we take...). When evenly distributing the pad surface , Mixed areas In this embodiment, the quantitative result of the underlying surface hybridity is used as a static driving factor for the grid resolution of NWP (Numerical Weather Prediction), which is used to drive the automatic sparsification of uniform regions and the automatic densification of heterogeneous regions.
[0051] Optionally, Shannon information entropy can be replaced by the Simpson diversity index or Rényi entropy, and type boundary density can be replaced by patch density, landscape fragmentation index, or aggregation index. These indicators can be used individually or in combination, ensuring that the resulting land use mix weights remain monotonically non-decreasing.
[0052] After obtaining the terrain complexity weight and land use mix weight, they are aggregated using an aggregation operator that maintains monotonicity. The aggregation operator satisfies the following condition: when either the terrain complexity weight or the land use mix weight increases while the other remains unchanged, the static resolution weight field remains unchanged or increases. In one embodiment, a maximum value aggregation operator is used.
[0053] This preserves the independent driving force of different static sub-weights and avoids the weakening of locally high-weight factors by the averaging process. Optionally, static sub-weights can also be weighted by summation, Aggregation is performed using a norm or a soft maximum operator. The soft maximum operator can take the form of log-sum-exp, but the aggregation operator used should maintain the monotonicity mentioned above.
[0054] The static resolution weight field can be pre-calculated before the model integration begins and reused during a single forecast integration process. The static resolution weight field can be regenerated when the digital elevation model data or land use and land cover type data is updated.
[0055] The dynamic driving layer calculates dynamic resolution weights based on the meteorological state field diagnosed in real time during the integration process of the meteorological numerical weather prediction model. This application defines it as a normalized aggregation of a set of meteorological diagnostic quantities related to high-impact weather processes; the types and number of indicators used can be increased or decreased, including but not limited to: convective available potential energy (CAPE), horizontal temperature gradient, etc. Water vapor flux divergence Wind shear Relative vorticity Lifting Index (LI / SI) These include exponents, frontogenesis function, potential vortex anomaly, vertical velocity, and precipitation rate. In one implementation, the maximum value after normalization of these five indicators is taken:
[0056] in, The trigger thresholds for each corresponding indicator are set, and the normalized values are limited to [a specific value]. (As an example, , Each trigger threshold can be configured according to the characteristics of the climate zone and the focus of the forecast, and the aggregation operator is not limited to taking the maximum value.
[0057] The aggregation operator for dynamic resolution weighted fields can also employ weighted summation. - Norm or soft maximum operator. Regardless of the aggregation method used, the dynamic resolution weight field remains unchanged or increases as any normalized meteorological diagnostic quantity increases. As the integration time of the meteorological numerical forecast model progresses and the weather system moves, develops, or dissipates, the dynamic resolution weight field forms a corresponding spatial distribution at different integration times.
[0058] The static driving layer and the dynamic driving layer are fused together by a nonlinear fusion function with positive synergistic (superlinear enhancement) properties. By coupling, the overall resolution control field is obtained. :
[0059] Nonlinear fusion function It satisfies the following essential characteristics: Monotonically non-decreasing for both driving layers: ,
[0060] It exhibits positive cross sensitivity (supermodel property / superlinear cooperation):
[0061] That is, the higher the value of any one driving layer, the greater the marginal encryption contribution of the other driving layer.
[0062] When complex underlying surface ( High) and severe weather processes ( When (high) overlaps in space and time, This achieves an additional synergistic improvement in resolution over the overlapping region, exceeding the sum of the independent contributions of the two layers (i.e., superlinear amplification).
[0063] In one embodiment, the nonlinear fusion function employs an additive-multiplicative fusion function:
[0064] in, For the linear gain coefficient of the dynamic driving layer, , For coupling enhancement coefficients (as an example, The third term in the above formula That is, the coupling enhancement term.
[0065] Other forms of nonlinear fusion functions can also be used: Power-law coherence: , ; Product amplification formula: ; Exponential Cooperative: .
[0066] In addition, learnable fusion models (such as monotonic neural networks) obtained by supervised learning from historical forecast errors and constrained to satisfy the above monotonicity and positive cross sensitivity.
[0067] In the above formula, The exponent parameter represents the product term of the static resolution weight field and the dynamic resolution weight field.
[0068] Overall resolution control field Then, through an arbitrary monotonically non-increasing mapping, it is converted into the target mesh edge length. In one implementation, a linear mapping is employed:
[0069] in, Indicates the target grid cell scale. Indicates the preset maximum grid cell size. This indicates the preset minimum grid cell size.
[0070] Mapping forms include, but are not limited to, linear, piecewise linear, power-law, exponential, and lookup table mappings. The larger The smaller the value, the more likely it is to fall within the scope of this application.
[0071] To ensure runtime efficiency, this application pre-constructs and keeps a multi-resolution hierarchical data structure resident in memory before mode startup, and decouples it from the resolution control field calculated in real time during runtime. This degrades mesh updates to table lookup and cell activation / sleep operations without requiring re-subdivision. The hierarchical data structure includes, but is not limited to: quadtrees, octrees, KD trees, hierarchical mesh trees with arbitrary refinement ratios (not limited to 2), and structured or unstructured multi-layer meshes; the mesh topology can be triangular, quadrilateral, hexagonal, cubic sphere, etc. During runtime, based on... Directly locate the target level corresponding to each spatial location and execute unit activation or sleep, with a single operation having constant or near-constant time complexity (example). In one implementation, taking level 0 as an example with a resolution of 9km and a refinement ratio of 2 per level, level 4 corresponds to approximately 0.6km, and the total number of levels is... It can be configured according to accuracy requirements.
[0072] Considering the evolutionary time characteristics of weather processes at different scales, this application designs a hierarchical (two-, three-, or more-level) triggering mechanism to match the update time interval of each grid level with the characteristic time scale of the corresponding weather process. When each level is triggered, only [the following applies]. The change relative to the previous time exceeds the corresponding control field change threshold. The computational grid cells are adjusted hierarchically to avoid the computational overhead of global re-partitioning; the control field change threshold increases with level ( Taking a three-level trigger as an example: Level 1 (coarse update) has a relatively long time interval (e.g., 6 hours), suitable for slow-changing processes at the scale of weather systems; Level 2 (medium update) is in the middle (e.g., 1 hour), suitable for hourly-scale changes such as mesoscale convective systems; Level 3 (fine update) has the shortest time interval (e.g., 15 minutes), suitable for rapidly evolving events such as severe convection. The number of levels and time intervals mentioned above are just examples and can be flexibly configured according to computing resources and business timeliness.
[0073] After each refinement or coarsening of the computational grid, a conservation remapping operator that satisfies variable integral conservation is used to remap the meteorological state variables constituting the meteorological state field on the original computational grid to the adjusted computational grid. Conservation remapping can employ first-order conservation remapping based on the area weighted by the geometric overlay of the old and new grids, second-order or higher-order conservation interpolation, spectral element conservation remapping, or remapping based on a supermesh. When computational resources are limited, first-order conservation remapping can be used.
[0074] Each time the grid level is adjusted, a conserved remapping operator that guarantees variable integral conservation is used to complete the variable transformation between the old and new grids. The conserved remapping includes, but is not limited to: first-order conserved remapping based on geometrically superimposed area weighting, second-order or higher-order conserved interpolation (such as spectral element conserved remapping, supermesh construction), etc.; when resources are limited, it can degenerate into first-order conserved interpolation. In one implementation, it is based on arbitrary old computational grid cells. With new computational grid cells The variable values for the new computational grid cells are:
[0075] in, This represents the variable values after remapping in the new computational grid cells. Indicates the first j The variable values of each old computational grid cell before remapping. Indicates the first j The old computing grid cell and the first i The geometric overlap area between the new computational grid cells, where Area represents the area calculation function. Indicates the first j The spatial region covered by an old computational grid cell.
[0076] This conservation remapping operator strictly satisfies global integral conservation (mass, momentum, and energy) and is applicable in both the refinement and coarsening directions of the computational grid. To further suppress spurious numerical noise caused by computational grid adjustment, this application sets a gradient resolution transition band around the refinement region and implements time-smooth soft switching during the computational grid adjustment process. Both the width of the transition band and the soft switching time constant are configurable.
[0077] After completing the conservation remapping, the meteorological numerical forecasting model continues to integrate the dynamic and physical processes on the updated computational grid, forming the meteorological state field for the next integration time. Upon reaching the next dynamic resolution weight field update time or the update time triggered by any computational grid update level, the meteorological state field is re-diagnosed and updated. ,calculate ,Sure And adjust the computation grid according to the corresponding triggering conditions.
[0078] In practical applications, when a weather system moves to a new spatial region, the high-value areas of the dynamic resolution weight field shift accordingly, enabling the high-resolution computational grid to dynamically adjust to follow the weather system. As the weather process weakens or moves out of its original area, the dynamic resolution weight at the corresponding location decreases, allowing the computational grid to revert to a lower resolution when coarsening conditions are met. In areas with complex terrain and mixed land use types, the static resolution weight field can maintain the required basic resolution. When a complex underlying surface overlaps with a severe weather process in space and time, the nonlinear fusion function generates synergistic enhancement, giving the overlapping area a higher priority in resolution configuration than if the two driving factors acted independently. In regions with a uniform underlying surface and stable weather conditions, both the static and dynamic resolution weight fields are at low values, allowing the computational grid to actively coarsen, thus concentrating computational resources on key areas affecting forecast accuracy.
[0079] As an example runtime configuration, a hierarchical grid data structure from level 0 to level 4 can be pre-built. Level 0 corresponds to a 9 km resolution, with a refinement factor of 2 between adjacent levels, and level 4 corresponds to approximately 0.6 km resolution. The computational grid updates employ a three-level triggering mechanism, with update intervals set to 6 hours, 1 hour, and 15 minutes, respectively, serving slowly varying weather system scales, mesoscale weather processes, and rapidly evolving strong convective processes. Two to three rings of gradually varying resolution transition cells can be set around the periphery of the local densification zone to reduce the impact of abrupt grid scale changes on model integral stability.
[0080] In scenarios involving severe convection, torrential rain, or typhoon landfall in mountainous areas, complex terrain results in higher static resolution weights for those regions. As the severe weather event moves into the area, the dynamic resolution weight increases synchronously, with both enhancing each other at the same spatiotemporal location. This further reduces the target grid cell size and activates higher resolution layers. As the severe weather event moves out of the area or gradually weakens, the dynamic resolution weight decreases accordingly. When coarsening conditions are met, the corresponding high-resolution grid cells can be put into dormancy, while complex terrain areas can still maintain a higher base grid resolution than uniform plain areas based on their higher static resolution weights.
[0081] In areas with fragmented land use and land cover types, such as urban-rural transition zones and agroforestry transition zones, the land use mixing weight can maintain the grid resolution required to express the spatial discontinuity of surface fluxes. In large areas of single underlying surface, such as plains, farmland, and uniform meadows, when the weather conditions are relatively stable, the computational grid can be coarsened to a lower resolution allowed by the system, thereby concentrating computational resources on complex underlying surfaces and areas with high weather impacts.
[0082] Compared to a global high-resolution grid, this application is expected to save approximately 30%–60% of computation. The dynamic triggering mechanism is expected to complete local refinement 2–3 forecast times in advance before strong convection occurs, and the short-term forecast skill score (TS) is expected to improve by approximately 10%–25%. With the adoption of a pre-computed hierarchical grid tree, the runtime grid adjustment time is expected to be controlled within seconds.
[0083] Optionally, the above method can also be adapted to different meteorological numerical forecasting models using triangular, quadrilateral, hexagonal, or cubic sphere grids, and can also be further used for ocean numerical forecasting, air quality simulation, or climate change scenario simulation. In the above applications, the appropriate dynamic state diagnostic quantities can be selected according to the simulation object, while maintaining the processing relationships between the static driving layer, dynamic driving layer, nonlinear collaborative fusion, target grid resolution mapping, and conserved grid adjustment.
[0084] like Figure 3 The diagram shown is a schematic representation of a four-dimensional spatiotemporal adaptive grid generation system for meteorological numerical weather prediction according to this application, which may include: The static resolution weight field module is used to obtain a static resolution weight field determined based on the underlying surface data of the forecast area. The static resolution weight field is used to characterize the static heterogeneity of the underlying surface of the forecast area. The dynamic resolution weight field module is used to obtain the meteorological state field of the forecast area at the current integration time during the integration process of the meteorological numerical forecast model, and to determine the dynamic resolution weight field based on the meteorological state field. The fusion module is used to fuse the static resolution weight field and the dynamic resolution weight field through a nonlinear fusion function to obtain a comprehensive resolution control field; wherein, the nonlinear fusion function satisfies the following: when either the static resolution weight field or the dynamic resolution weight field increases while the other remains unchanged, the comprehensive resolution control field remains unchanged or increases; when both the static resolution weight field and the dynamic resolution weight field increase together, the increment of the comprehensive resolution control field is greater than the sum of the increments produced by the individual increases of the static resolution weight field and the dynamic resolution weight field. The target grid resolution field module is used to determine the target grid resolution field based on the comprehensive resolution control field. The update module is used to refine or coarse the computational grid of the prediction region according to the target grid resolution field, so as to obtain a four-dimensional spatiotemporal adaptive grid that updates with the integration time.
[0085] It should be noted that, in the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of each block is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple blocks may be combined or integrated into another device, or some features may be ignored or not executed. The modules described as separate components may or may not be physically separated. The components shown as modules may be one or more physical units, that is, they may be located in one place or distributed in multiple different places. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs.
[0086] Furthermore, in the various embodiments of the present invention, the modules can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The integrated unit described above can be implemented in hardware or as a software functional unit.
[0087] This application also provides an electronic device, which may include one or more processors, memory and communication interfaces.
[0088] The memory, communication interface, and processor are coupled together. For example, the memory, communication interface, and processor can be coupled together via a bus.
[0089] The communication interface is used for data transmission with other devices. The memory stores computer program code. This computer program code includes computer instructions, which, when executed by the processor, cause the electronic device to perform the steps of the aforementioned four-dimensional spatiotemporal adaptive grid generation method for meteorological numerical forecasting.
[0090] The processor can be a processor or controller, such as a Central Processing Unit (CPU), a general-purpose processor, a Digital Signal Processor (DSP), an Application-Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with this disclosure. The processor can also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc. The processor can be used to support an electronic device in performing the method steps provided in the above embodiments.
[0091] The bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. These buses can be categorized as address buses, data buses, control buses, etc.
[0092] This application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of the above-described four-dimensional spatiotemporal adaptive grid generation method for meteorological numerical forecasting.
[0093] The computer-readable storage media involved in this application include random access memory (RAM), memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage media known in the art.
[0094] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A method for generating a four-dimensional spatio-temporal adaptive mesh for numerical weather prediction, characterized in that, include: Obtain a static resolution weight field determined based on the underlying surface data of the forecast area, wherein the static resolution weight field is used to characterize the static heterogeneity of the underlying surface of the forecast area; During the integration process of the meteorological numerical forecasting model, the meteorological state field of the forecast area at the current integration time is obtained, and the dynamic resolution weight field is determined based on the meteorological state field. The static resolution weight field and the dynamic resolution weight field are fused using a nonlinear fusion function to obtain a comprehensive resolution control field. The nonlinear fusion function satisfies the following: when one term in the static resolution weight field or the dynamic resolution weight field increases while the other term remains unchanged, the comprehensive resolution control field remains unchanged or increases; when both the static resolution weight field and the dynamic resolution weight field increase together, the increment of the comprehensive resolution control field is greater than the sum of the increments produced by the individual increases of the static resolution weight field and the dynamic resolution weight field. The nonlinear fusion function can be an additive-multiplicative fusion function, a power-law cooperative fusion function, a product-amplified fusion function, an exponential cooperative fusion function, or a learnable fusion model. The target grid resolution field is determined based on the comprehensive resolution control field. The computational grid of the prediction region is refined or coarsened according to the target grid resolution field to obtain a four-dimensional spatiotemporal adaptive grid that is updated with the integration time. 2.The method according to claim 1, wherein, The method for determining the static resolution weight field includes: Obtain digital elevation model data and land use and land cover type data for the forecast area; The terrain complexity weight is determined based on the digital elevation model data, and the land use mixing weight is determined based on the land use and land cover type data. The terrain complexity weights and land use hybridity weights are aggregated using a monotonic aggregation operator to obtain the static resolution weight field.
3. The method of claim 2, wherein, The aggregation operator that maintains monotonicity satisfies the following condition: when either the terrain complexity weight or the land use hybridity weight increases, the static resolution weight field remains unchanged or increases. 4.The method according to claim 1, wherein, Determining the dynamic resolution weight field based on the meteorological state field includes: Based on the meteorological state field, at least one meteorological diagnostic quantity related to high-impact weather processes is determined; The meteorological diagnostic data are normalized. The normalized meteorological diagnostic data are aggregated using a monotonic aggregation operator to obtain the dynamic resolution weight field.
5. The method of claim 1, wherein, The nonlinear fusion function is a learnable fusion model; The learnable fusion model is trained under supervision based on historical weather forecast errors, and monotonicity constraints and collaborative enhancement constraints are applied during the training process to ensure that the learnable fusion model satisfies the conditions of the nonlinear fusion function.
6. The four-dimensional spatiotemporal adaptive grid generation method for meteorological numerical forecasting according to claim 1, characterized in that, The target grid resolution field is composed of the target grid cell scale corresponding to each spatial location in the prediction area; Determining the target grid resolution field based on the comprehensive resolution control field includes: The integrated resolution control field is converted into the target grid cell scale corresponding to each spatial location in the forecast area by a monotonically non-increasing mapping, so that when the integrated resolution control field increases, the corresponding target grid cell scale remains unchanged or decreases.
7. The four-dimensional spatiotemporal adaptive grid generation method for meteorological numerical forecasting according to claim 1, characterized in that, The computational grid is formed based on a pre-built and stored hierarchical grid data structure, which includes multiple grid resolution levels. The computational grid of the prediction region is refined or coarsened according to the target grid resolution field, including: The target grid level corresponding to each spatial location is determined based on the target grid resolution field, and the computational grid is refined or coarsened by activating or suspending the computational grid cells corresponding to the grid resolution level.
8. The four-dimensional spatiotemporal adaptive grid generation method for meteorological numerical forecasting according to claim 7, characterized in that, When refining or coarsening the computational grid by activating or suspending computational grid units at the corresponding grid resolution level, the process is executed according to multiple grid update trigger levels. The multiple grid update trigger levels correspond to the characteristic time scales of weather processes of different scales, and each has a corresponding grid update time interval and control field change threshold. When the grid update time corresponding to any grid update trigger level arrives: Determine the amount of change in the integrated resolution control field relative to the previous grid update time; For computational grid cells whose change exceeds the corresponding control field change threshold, the corresponding target grid level is determined based on the target grid resolution field, and the computational grid cells at the corresponding grid resolution level are activated or dormant.
9. The four-dimensional spatiotemporal adaptive grid generation method for meteorological numerical forecasting according to claim 1, characterized in that, After refining or coarsening the computational grid of the forecast area according to the target grid resolution field, the method further includes: using a conservation remapping operator that satisfies variable integral conservation to remap the meteorological state variables that constitute the meteorological state field on the computational grid before refinement or coarsening to the computational grid after refinement or coarsening.
10. A four-dimensional spatiotemporal adaptive grid generation system for meteorological numerical forecasting, characterized in that, include: The static resolution weight field module is used to obtain a static resolution weight field determined based on the underlying surface data of the forecast area. The static resolution weight field is used to characterize the static heterogeneity of the underlying surface of the forecast area. The dynamic resolution weight field module is used to obtain the meteorological state field of the forecast area at the current integration time during the integration process of the meteorological numerical forecast model, and to determine the dynamic resolution weight field based on the meteorological state field. A fusion module is used to fuse the static resolution weight field and the dynamic resolution weight field using a nonlinear fusion function to obtain a comprehensive resolution control field. The nonlinear fusion function satisfies the following: when one term in the static resolution weight field and the dynamic resolution weight field increases while the other term remains unchanged, the comprehensive resolution control field remains unchanged or increases; when both the static resolution weight field and the dynamic resolution weight field increase together, the increment of the comprehensive resolution control field is greater than the sum of the increments produced by the individual increases of the static resolution weight field and the dynamic resolution weight field. The nonlinear fusion function can be an additive-multiplicative fusion function, a power-law cooperative fusion function, a product-amplified fusion function, an exponential cooperative fusion function, or a learnable fusion model. The target grid resolution field module is used to determine the target grid resolution field based on the comprehensive resolution control field. The update module is used to refine or coarse the computational grid of the prediction region according to the target grid resolution field, so as to obtain a four-dimensional spatiotemporal adaptive grid that is updated with the integration time.
11. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the four-dimensional spatiotemporal adaptive grid generation method for meteorological numerical forecasting as described in any one of claims 1-9.
12. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the four-dimensional spatiotemporal adaptive grid generation method for meteorological numerical forecasting as described in any one of claims 1-9.
Citation Information
Patent Citations
Meteorological element forecasting method based on graph neural network multi-modal meteorological data fusion
CN116720156A
Meteorological big data intelligent diagnosis and early warning system and method
CN122173900A
