An ai large model-based multi-temporal and spatial runoff prediction method and system

Abstract

The application relates to the technical field of runoff prediction, and provides a multi-time-space runoff prediction method and system based on an AI large model, which comprises the following steps: constructing a three-dimensional geological model by combining drilling data with digital elevation data, converting the three-dimensional geological model into a hydraulic conduction connected graph with a main permeation direction after density self-adaptive grid division, and accurately depicting hydrological connectivity; meanwhile, terrain features such as a surface ditch line are extracted, underground strong permeation channels are matched, and underground water supply marks are given, so that the underground and surface hydrological processes are coupled; finally, a graph attention neural network is trained by using rainfall and runoff data, a hydraulic offset value positively correlated with a confluence cumulative value is added during the training, a feature propagation is guided by using a physical mechanism, and the runoff prediction precision under a complex terrain is improved.

Description

Technical Field

[0001] This invention relates to the field of runoff prediction technology, and more specifically, to a multi-temporal and spatial runoff prediction method and system based on an AI large model. Background Technology

[0002] The content in this section only provides background information related to this invention and may not constitute prior art.

[0003] Runoff prediction is one of the core research directions in the field of hydrology and water resources, and its results directly support key applications such as flood control and disaster reduction, optimal allocation of water resources, and safe operation of water conservancy projects. Traditional hydrophysical models rely on a large number of prior parameters and complex physical equations, which limits their applicability in areas with scarce data or complex geological conditions. Although early data-driven models can learn hydrological patterns from historical data, they generally lack effective characterization of spatial topological relationships and hydrophysical mechanisms. In recent years, runoff prediction methods based on graph neural networks have become a research hotspot because they can naturally capture watershed spatial dependencies, but they still have significant shortcomings in the coupled simulation of groundwater hydrological processes and surface runoff.

[0004] In existing technologies, such as the Chinese patent with publication number CN120105890A, a rainfall-runoff simulation method integrating knowledge graphs and graph attention is disclosed. This method uses a graph attention neural network as the core model, abstracting hydrological stations and sub-basins as graph nodes and defining edge weights based on upstream and downstream topological relationships. However, this method does not utilize drilling data to construct a three-dimensional geological model, thus failing to characterize the heterogeneity and permeability of different underground geological layers, and also failing to consider the recharge effect of strong underground permeability channels on surface runoff, resulting in a significant decrease in prediction accuracy in areas with significant groundwater recharge. Furthermore, its training process does not incorporate constraints based on hydrophysical mechanisms, and feature propagation cannot effectively favor areas with strong surface runoff, making it difficult to accurately simulate complex hydrological processes.

[0005] For example, Chinese patent CN115511166A discloses a multi-step spatiotemporal prediction method for watershed-scale runoff based on a combination of static and dynamic graphs. This method uses digital elevation data to extract topographic features and constructs a graph structure that integrates spatial topological relationships to achieve multi-spatiotemporal scale prediction. However, the graph structure constructed by this method is only based on river channel distance and elevation difference, lacking an accurate representation of surface hydrological connectivity, and it does not incorporate a physical guidance mechanism in model training, making it difficult to adapt to watershed scenarios with complex and varied terrain.

[0006] Therefore, there is an urgent need for a runoff prediction method to address the problems of existing technologies, such as the lack of accurate characterization of surface hydrological connectivity, insufficient integration of topographic runoff characteristics, and low prediction accuracy under complex terrain conditions due to the lack of physical mechanisms to guide the prediction. Summary of the Invention

[0007] The purpose of this invention is to provide a multi-temporal and spatial runoff prediction method and system based on a large AI model to improve the aforementioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows: Firstly, this application provides a multi-temporal and spatial runoff prediction method based on a large AI model, including: Acquire drilling data and digital elevation data for the target area; A geological model of the target area is constructed using drilling data. The drilling data is obtained by setting up multiple sampling points in the target area and drilling and core sampling at each sampling point to obtain geological stratification data and porosity data for each sampling point. Based on the geological stratification data and the corresponding digital elevation data, a three-dimensional geological model of the target area is established. Based on the sampling density of drilling data, the 3D geological model is divided into density-adaptive grids. Each geological layer within each grid cell is converted into a graph node containing a main permeability direction vector, which is calculated based on the effective porosity and average pore size of the geological layer. Vertical edges between nodes of different geological layers within the same grid cell, as well as horizontal edges between adjacent nodes, are constructed. The weight of any edge is obtained by weighted calculation based on the effective porosity and average pore size of different geological layers in the corresponding directions of two adjacent nodes, forming a hydraulic conduction connectivity graph of the target area. Simultaneously, based on digital elevation data, surface depressions, valley lines, and the cumulative confluence values ​​of valley lines are extracted. Nodes of underground strong permeability channels with permeability coefficients exceeding a preset threshold in the hydraulic conduction connectivity graph are spatially projected and matched with valley lines, assigning groundwater recharge identifiers to the matched valley line segments. Acquire rainfall distribution data and historical surface runoff observation data for the target area; map the rainfall distribution data into initial features of the corresponding surface nodes in the hydraulic conduction connectivity graph; convert the historical surface runoff observation data into runoff ground truth labels for the observation point locations; and extract the topographic boundaries of unobserved areas as auxiliary constraints to train the graph attention neural network; during training, in the forward propagation, force the edge weights from the nodes of strong underground infiltration channels to the surface nodes corresponding to the valley lines that are given groundwater recharge indicators to add a hydraulic offset that is positively correlated with the cumulative runoff value, so that the feature propagation is biased towards the strong surface runoff areas, thereby obtaining the runoff prediction model; Input real-time rainfall distribution data into the runoff prediction model and output prediction results that include the location of surface runoff at multiple temporal and spatial scales.

[0008] Furthermore, the lowest point of the sampling depth at each sampling point is the lowest point of the nearest surface depression or valley line.

[0009] Furthermore, based on geological stratification data and corresponding digital elevation data, a three-dimensional geological model of the target area is established, specifically including: Based on the lithological type and sedimentary sequence characteristics of the geological strata, the geological strata of all sampling points are uniformly numbered and stratigraphically correlated to establish a stratigraphic framework for the target area; Under the constraints of the stratigraphic framework, spatial interpolation is performed on the elevation of the top and bottom interfaces of the same geological stratum between adjacent sampling points to generate continuous geological strata. The geological strata are smoothed to eliminate unreasonable fluctuations caused by interpolation; the continuous geological strata are then integrated with the surface model generated from digital elevation data to form an initial three-dimensional geological model of the target area. The initial three-dimensional geological model is topologically checked to correct errors such as overlapping, intersecting, and missing geological layers, thus obtaining a three-dimensional geological model of the target area.

[0010] Furthermore, the density-adaptive mesh generation process also includes: Calculate the average drilling sampling interval for the target area; If the distance between any two sampling points within any grid cell is less than half of the average drilling sampling interval, then the grid cell is recursively subdivided; the subdivision stops when the distance between any two sampling points within all grid cells is not less than half of the average drilling sampling interval.

[0011] Furthermore, each geological layer within each grid cell is converted into a graph node containing a principal permeation direction vector, specifically including: For each geological layer within each grid cell, the effective porosity and average pore size of the corresponding geological layer in the horizontal and vertical directions are obtained from the pore data; the vertical direction is the direction of gravity, and the horizontal direction is the direction perpendicular to gravity. The permeability intensity in the corresponding direction is obtained by multiplying the effective porosity and the average pore size in the corresponding direction; the main permeability direction vector of each geological layer is obtained by vector merging based on the permeability intensity in different directions. Each geological stratum within each grid cell is treated as an independent unit, and the main permeation direction vector is used as the core attribute of that unit to construct graph nodes.

[0012] Furthermore, the weight of any edge is obtained by weighting the effective porosity and average pore size in different geological layers in the corresponding directions of two adjacent nodes; For the vertical edge, obtain the effective porosity and average pore size of the two adjacent nodes in the vertical direction respectively, calculate the vertical permeability of the two graph nodes respectively, and use the thickness ratio of the two geological layers in the grid cell as the weight to perform a weighted average of the two vertical permeability intensities to obtain the basic weight of the vertical edge. For a horizontal edge, identify the geological stratification groups located in the same horizontal direction between two adjacent nodes; for each geological stratification group, obtain the effective porosity and average pore size in the horizontal direction of the two nodes respectively, and calculate the horizontal permeability intensity of each group; use the length ratio of the geological stratification group between the two nodes as the weight, and perform a weighted summation of the average values ​​of each group to obtain the weight of the horizontal edge.

[0013] Furthermore, based on digital elevation data, the confluence and cumulative values ​​of surface depressions, valley lines, and valley lines are extracted, specifically including: A terrain-filling algorithm is used to fill local depressions in digital elevation data to generate a depression-free digital elevation model. The water flow direction of each grid cell is calculated based on the depression-free digital elevation model. The cumulative runoff value of each grid cell is calculated according to the water flow direction. Grid cells with cumulative runoff values ​​exceeding a preset threshold are connected to form valley lines. Unfilled permanent depressions in the digital elevation data are identified and marked as surface depressions.

[0014] Furthermore, the terrain boundaries of unobserved areas are extracted as auxiliary constraints, specifically including: Identify the areas within the target region where no surface runoff observation points have been set up; extract the ridgelines, watersheds, and abrupt topographic changes in the unobserved areas as topographic boundaries; convert the topographic boundaries into constraint edges in the hydraulic conduction connectivity graph; during the training of the graph attention neural network, restrict the propagation direction of features on the constraint edges, so that features can only propagate from high-altitude areas to low-altitude areas; at the same time, restrict the propagation intensity of features on the constraint edges, so that the feature propagation intensity across topographic boundaries is lower than the feature propagation intensity within the same topographic unit.

[0015] Secondly, this application also provides a multi-temporal runoff prediction system based on an AI large model, including: The data acquisition module is used to acquire drilling data and digital elevation data for the target area; The geological modeling module is used to construct a geological model of the target area using drilling data. The drilling data is obtained by setting up multiple sampling points in the target area, drilling and core sampling at each sampling point to obtain geological stratification data and porosity data for each sampling point; based on the geological stratification data and the corresponding digital elevation data, a three-dimensional geological model of the target area is established. The connectivity graph construction module is used to perform density-adaptive meshing of the 3D geological model based on the sampling density of drilling data. Each geological layer within each grid cell is converted into a graph node containing a main permeability direction vector, which is calculated based on the effective porosity and average pore size of the geological layer. Vertical edges between nodes of different geological layers within the same grid cell, as well as horizontal edges between adjacent nodes, are constructed. The weight of any edge is obtained by weighted calculation based on the effective porosity and average pore size of different geological layers in the corresponding directions of two adjacent nodes, forming a hydraulic conduction connectivity graph of the target area. Simultaneously, based on digital elevation data, surface depressions, valley lines, and the cumulative confluence values ​​of valley lines are extracted. The nodes of underground strong permeability channels with permeability coefficients exceeding a preset threshold in the hydraulic conduction connectivity graph are spatially projected and matched with the valley lines, and groundwater recharge identifiers are assigned to the matched valley line segments. The model training module is used to acquire rainfall distribution data and historical surface runoff observation data of the target area; it maps the rainfall distribution data into the initial features of the corresponding surface nodes in the hydraulic conduction connectivity graph, converts the historical surface runoff observation data into runoff ground truth labels of the observation point locations, and extracts the topographic boundaries of unobserved areas as auxiliary constraints to train the graph attention neural network; during training, in the forward propagation, the edge weights from the nodes of strong underground infiltration channels to the corresponding surface nodes of the valley lines that are given groundwater recharge indicators are forced to add a hydraulic offset that is positively correlated with the cumulative runoff value, so that the feature propagation is biased towards the strong surface runoff areas, thereby obtaining the runoff prediction model; The results module is used to input real-time rainfall distribution data into the runoff prediction model and output prediction results that include the location of surface runoff at multiple temporal and spatial scales.

[0016] Thirdly, this application also provides an electronic device, including: Memory, used to store computer programs; A processor for implementing the method steps as described in the first aspect when executing the computer program.

[0017] The beneficial effects of this invention are as follows: This invention obtains geological stratification and porosity data of the target area through drilling, and constructs a three-dimensional geological model by combining it with digital elevation data. Then, based on the drilling sampling density, a density-adaptive grid is generated, transforming the geological strata within each grid cell into graph nodes with a main permeability direction. Edges representing vertical connectivity between different geological strata and horizontal connectivity between adjacent nodes are constructed. Edge weights are calculated based on the porosity characteristics of adjacent geological strata to form a hydraulic conductivity connectivity graph. Utilizing the inherent suitability of graph structures for representing connectivity relationships between nodes, this invention accurately depicts the differences in hydraulic conductivity between different geological strata and at different locations within the same stratum, solving the core problem of traditional methods' inability to accurately represent hydrological connectivity under complex geological conditions. Simultaneously, based on digital elevation data, surface depressions, valley lines, and the cumulative runoff values ​​of valley lines are extracted. Underground strong permeability channel nodes are spatially projected and matched with valley lines, and groundwater recharge indicators are assigned. The physical laws governing the main paths and convergence areas of surface runoff are determined by topographic runoff characteristics, achieving deep coupling between underground and surface hydrological processes. This fully integrates the control of topography over the runoff process, overcoming the shortcomings of existing technologies that sever the connection between underground and surface hydrology. Based on this, rainfall distribution is mapped to the initial features of surface nodes corresponding to the hydraulic conduction connectivity graph. Historical surface runoff observation data is used as labels, and the topographic boundaries of unobserved areas are used as auxiliary constraints to train the graph attention neural network. During the forward propagation process, the edge weights of the nodes from the strong underground infiltration channels to the surface nodes corresponding to the valley lines with groundwater recharge indicators are forced to increase the hydraulic offset positively correlated with the cumulative runoff value. This physical constraint guides the feature propagation direction of the model to favor areas with strong surface runoff, so that the learning process of the artificial intelligence model strictly follows the inherent physical mechanism of the hydrological cycle, rather than simply relying on data fitting. Ultimately, this improves the physical rationality and generalization ability of runoff prediction under complex terrain conditions, and achieves accurate prediction of surface runoff location at multiple spatiotemporal scales. Attached Figure Description

[0018] Figure 1 A flowchart of a multi-temporal and spatial runoff prediction method based on an AI large model provided by the present invention; Figure 2 A schematic diagram of a multi-temporal runoff prediction system based on an AI large model provided by the present invention; Figure 3 This is a schematic diagram of an electronic device provided by the present invention.

[0019] In the diagram: 201, Data Acquisition Module; 202, Geological Modeling Module; 203, Connectivity Graph Construction Module; 204, Model Training Module; 205, Results Module; 301, Processor; 302, Memory. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] like Figure 1 As shown in the embodiment of the present invention, a multi-temporal and spatial runoff prediction method based on an AI large model includes: S101, acquire drilling data and digital elevation data for the target area.

[0022] Specifically, obtaining drilling data is based on the principle that the underground geological structure and pore characteristics directly determine the storage and transport capacity of groundwater, and core drilling is the most direct and reliable means of obtaining real underground geological information. Through drilling data, the lithological distribution and pore development at different depths can be accurately grasped, avoiding the estimation errors of groundwater hydrological parameters caused by relying solely on surface data, and significantly improving the accuracy of subsequent hydraulic conduction connectivity map construction.

[0023] Simultaneously acquiring digital elevation data of the target area is based on the principle that surface topography is a core factor controlling the formation, confluence direction, and confluence velocity of surface runoff. Digital elevation data can accurately express the elevation undulation characteristics of the surface in raster form. This data is not only used to construct the fusion of surface models and three-dimensional geological models, but also provides a foundation for subsequent extraction of surface hydrological elements such as surface depressions, valley lines, and runoff accumulation values, ensuring that the spatial coupling relationship between surface runoff processes and groundwater transport processes can be accurately depicted.

[0024] S102, Construct a geological model of the target area using drilling data. The drilling data is obtained by setting up multiple sampling points in the target area and drilling and core sampling at each sampling point to obtain geological stratification data and pore data for each sampling point. Based on the geological stratification data and the corresponding digital elevation data, a three-dimensional geological model of the target area is established.

[0025] Specifically, drilling data is obtained by setting up multiple sampling points in the target area and performing core drilling at each point to acquire geological stratification and porosity data. Core drilling, as a direct means of obtaining subsurface geological information, can accurately identify lithological types, stratigraphic thickness, and porosity development characteristics at different depths. These parameters directly determine the storage and transport capacity of groundwater, offering higher accuracy and reliability compared to indirect geophysical methods. The lowest sampling depth at each sampling point is the lowest point of the nearest surface depression or valley line. This design is based on the principle that surface depressions and valley lines are the main interfaces for groundwater and surface water exchange. Extending the sampling depth to this location allows for the complete acquisition of the geological strata involved in the interaction between groundwater and surface water, avoiding the loss of deep hydrogeological information due to insufficient sampling depth and ensuring the integrity of subsequent surface-groundwater hydrological process coupling analysis. For example, in a mountainous watershed, if the lowest point of the valley line closest to a sampling point is at an elevation of 20 meters, then the drilling depth of that sampling point needs to reach the depth corresponding to an elevation of 20 meters in order to fully cover all geological layers in the area that may be involved in the hydrological cycle.

[0026] Subsequently, based on geological stratification data and corresponding digital elevation data, a three-dimensional geological model of the target area is established. Digital elevation data provides an accurate surface elevation benchmark, which can convert the relative depth of underground geological stratification into absolute elevation, achieving spatial unification between the underground geological model and the surface topographic model.

[0027] Specifically, firstly, based on the lithological types and sedimentary sequence characteristics of geological strata, all sampling points are uniformly numbered and stratigraphically correlated to establish a stratigraphic framework for the target area. Sedimentary sequence characteristics reflect the evolutionary patterns of sedimentary environments throughout geological history. Strata formed in the same period share similar lithological characteristics and sedimentary sequences. Stratigraphic correlation can identify stratigraphic units belonging to the same geological period in different sampling points, establishing a regional-scale spatial correspondence between stratigraphic units. This eliminates inconsistencies in geological strata numbering across different sampling points, providing a unified stratigraphic constraint for subsequent spatial interpolation and preventing erroneous connections between strata from different periods.

[0028] Next, under the constraints of the stratigraphic framework, spatial interpolation is performed on the elevations of the top and bottom interfaces of the same geological stratum between adjacent sampling points to generate continuous geological bedding planes. The spatial interpolation method utilizes the spatial continuity of geological bodies to estimate the interface elevations of unknown areas using the known stratigraphic interface elevations of sampling points. The constraints of the stratigraphic framework ensure that the interpolation results conform to geological evolution patterns, avoiding interface intersections that do not conform to the depositional sequence, and significantly improving the spatial rationality of geological bedding planes.

[0029] Spatial interpolation employs an inverse distance weighted method. Based on the assumption of spatial continuity of the geological body, sampling points closer to the interpolation point have greater weights. The stratigraphic interface elevation of unknown points is calculated through weighted averaging, while avoiding interface intersections due to stratigraphic framework constraints. The formula is as follows:

[0030] In the formula, coordinates First Interpolated elevation of the top surface of each geological stratum; The number of neighboring sampling points participating in the interpolation; coordinates With the Euclidean distance between neighboring sampling points; The distance decay power exponent; For the first The absolute elevation of the top surface of the kth geological stratum at each sampling point.

[0031] Then, the geological strata are smoothed to eliminate unreasonable fluctuations caused by interpolation. During spatial interpolation, sharp fluctuations that do not conform to actual geological conditions may occur due to local variations in the sampling point data. Smoothing preserves the overall trend characteristics of the geological strata while eliminating local anomalies, making the model more consistent with the morphology of real geological bodies. The continuous geological strata are then fused with the surface model generated from digital elevation data to form the initial three-dimensional geological model of the target area. This achieves accurate spatial matching between subsurface geological structures and surface topography, laying the foundation for subsequent coupled analysis of surface-subsurface hydrological processes.

[0032] The smoothing process involves taking the arithmetic mean of the elevation values ​​of each interpolated raster cell and its eight adjacent cells. This eliminates sharp local fluctuations caused by spatial interpolation while preserving the overall trend of geological strata. The formula is as follows:

[0033] In the formula, Smoothed coordinates First The top surface elevation of each geological stratum; , The raster resolution of digital elevation data; This is the raster offset index in the x-direction; This is the raster offset index in the y-direction.

[0034] Finally, a topological check was performed on the initial 3D geological model to correct errors such as overlapping, intersecting, and missing geological layers, resulting in a 3D geological model of the target area. The topological check identifies potential geometric errors during model construction, which can lead to the failure of subsequent hydraulic conductivity connectivity diagram construction or incorrect hydrological parameter calculations. By correcting these errors, the topological consistency and geometric correctness of the 3D geological model can be ensured, providing a reliable spatial data foundation for subsequent hydrological analysis.

[0035] S103. Based on the sampling density of drilling data, the three-dimensional geological model is divided into density-adaptive grids. Each geological layer within each grid cell is converted into a graph node containing a main permeability direction vector, which is calculated based on the effective porosity and average pore size of the geological layer. Vertical edges between nodes of different geological layers within the same grid cell, as well as horizontal edges between adjacent nodes, are constructed. The weight of any edge is calculated by weighting the effective porosity and average pore size of different geological layers in the corresponding directions of two adjacent nodes, forming a hydraulic conduction connectivity graph of the target area. Simultaneously, based on digital elevation data, surface depressions, valley lines, and the cumulative confluence values ​​of valley lines are extracted. The nodes of underground strong permeability channels with permeability coefficients exceeding a preset threshold in the hydraulic conduction connectivity graph are spatially projected and matched with the valley lines, and the matched valley line segments are assigned groundwater recharge identifiers. Specifically, based on the sampling density of drilling data, a density-adaptive mesh is applied to the 3D geological model. The principle is that the spatial distribution of drilling sampling points directly determines the reliability of geological information. Densely sampled areas have high-precision geological data, requiring finer meshes to preserve local geological features; sparsely sampled areas have high-uncertainty geological data, and overly fine meshes introduce significant interpolation errors. This invention employs a recursive subdivision strategy to achieve density-adaptive meshing: the average drilling sampling interval of the target area is calculated, reflecting the overall sampling density level of the geological data in the entire area; if the distance between any two sampling points within any mesh cell is less than half the average drilling sampling interval, then that mesh cell is recursively subdivided. This design automatically generates finer meshes in densely sampled areas, ensuring that changes in geological information within each mesh cell are accurately captured. Mesh subdivision stops when the distance between any two sampling points within all mesh cells is no less than half the average drilling sampling interval. At this point, the resolution of all mesh cells matches the reliability of the local geological data, avoiding both the loss of geological features caused by overly coarse meshes and the surge in computational load and amplification of interpolation errors caused by overly fine meshes. For example, if the average sampling interval during drilling in a certain watershed is 1000 meters, and there are two sampling points in an initial grid cell with a distance of only 400 meters (less than 500 meters), then the grid cell is divided into four equal parts; if there are still sampling points with a distance of less than 500 meters in the sub-grids after subdivision, then the grid is divided into four equal parts again until all sub-grids meet the condition.

[0036] The formula for calculating the sampling interval is as follows:

[0037] In the formula, The average sampling interval for drilling in the target area; For the first The sampling point and the first Euclidean distance between sampling points; This represents the total number of drilling sampling points within the target area.

[0038] After mesh generation, each geological stratum within each grid cell is converted into a graph node containing a principal permeability direction vector. This vector is calculated based on the effective porosity and average pore size of the geological stratum. The principle is that groundwater migration in porous media exhibits significant anisotropy, with substantial differences in permeability across different directions. Effective porosity and average pore size are the core parameters determining permeability. The specific implementation process is as follows: For each geological stratum within each grid cell, the effective porosity and average pore size of the corresponding geological stratum in the horizontal and vertical directions are obtained from the pore data. The vertical direction is the gravity direction, and the horizontal direction is the direction perpendicular to gravity. This directional definition conforms to the basic physical laws of groundwater transport. Gravity is the main driving force for vertical groundwater transport, while horizontal transport is mainly controlled by the hydraulic gradient. The permeability intensity in the corresponding direction is obtained by multiplying the effective porosity and average pore size in the corresponding direction. This product can comprehensively reflect the connectivity and water flow capacity of the pore space. The effective porosity determines the proportion of pore volume involved in groundwater transport, and the average pore size determines the water flow capacity of the pore channels. Based on the permeability intensity in different directions, vector merging is performed to obtain the main permeability direction vector of each geological stratum. This vector can intuitively express the dominant transport direction and intensity of groundwater within that geological stratum. Each geological stratum within each grid cell is taken as an independent unit, and the main permeability direction vector is used as the core attribute of the unit to construct graph nodes. This node construction method can accurately characterize the anisotropic features of groundwater transport, significantly improving the accuracy of groundwater transport process simulation compared to the traditional homogeneous permeability coefficient assumption.

[0039] The formulas for calculating the permeability intensity in each direction are as follows:

[0040] In the formula, For the first Within the first grid cell Horizontal permeability of each geological layer For the index of the grid cell; For the first At the sampling point, the first The effective porosity of each geological stratum; For the first At the sampling point, the first Average size of horizontal pores in each geological stratum.

[0041]

[0042] In the formula, For the first Within the first grid cell Vertical permeability of each geological stratum; For the first At the sampling point, the first Vertical effective porosity of each geological stratum; For the first At the sampling point, the first Average size of vertical pores in each geological stratum.

[0043] The formula for combining the main penetration direction vector is:

[0044] In the formula, For the first Within the first grid cell The main permeability direction vector of each geological stratum; It is a unit vector in the horizontal direction; It is a unit vector in the vertical direction.

[0045] Next, vertical edges between nodes of different geological strata within the same grid cell, as well as horizontal edges between adjacent nodes, are constructed. The weight of any edge is obtained by weighted calculation based on the effective porosity and average pore size within different geological strata in the corresponding directions of two adjacent nodes, forming a hydraulic conduction connectivity graph of the target area. The principle is that the edges in the graph structure represent the migration channels of groundwater between different nodes, and the edge weight reflects the hydraulic conduction capacity of the channel; a larger weight indicates that groundwater is more easily transported through the channel. For vertical edges, they represent the vertical exchange process of groundwater between adjacent geological strata within the same grid cell. The effective porosity and average pore size in the corresponding vertical direction of two adjacent nodes are obtained, and the vertical permeability of the two graph nodes is calculated separately. Using the thickness proportion of the two geological strata within the grid cell as weights, the two vertical permeability intensities are weighted and averaged to obtain the basic weight of the vertical edge. The thickness proportion weighting accurately reflects the contribution of different geological strata to the vertical hydraulic conduction capacity; thicker geological strata have a greater impact on the overall vertical conduction capacity. For a horizontal edge, it represents the horizontal transport process of groundwater between the same geological strata within adjacent grid cells. The geological strata groups located in the same horizontal direction between two adjacent nodes are identified. For each geological stratum corresponding to a group, the effective porosity and average pore size in the horizontal direction of the two nodes are obtained, and their respective horizontal permeability is calculated. The weight of the horizontal edge is obtained by weighting the average values ​​of each group based on the proportion of the length of the corresponding geological strata between the two nodes. The length proportion weighting accurately reflects the contribution of different geological strata segments to the horizontal hydraulic conduction capacity; longer geological strata segments have a greater impact on the overall horizontal conduction capacity. This edge weighting method accurately quantifies the hydraulic conduction capacity between different directions and different geological strata, enabling the hydraulic conduction connectivity diagram to truly reflect the actual transport patterns of groundwater flow.

[0046] Among them, for the vertical edge weight, the thickness ratio of two adjacent geological layers is used as the weight, and their respective vertical permeability is weighted and averaged to accurately reflect the greater contribution of thicker geological layers to the vertical hydraulic conduction capacity. Its expression is:

[0047] In the formula, For the first Within the first grid cell The and the first +1 vertical edge weight between geological layers; , The first At the sampling point, the first The, the +1 geological layer thickness; , The first Within the first grid cell The, the +1 vertical permeability of geological strata; For the horizontal edge weight, the average horizontal permeability intensity of the corresponding segments of two grid cells is taken for each geological layer, and then a weighted sum is performed using the length ratio of each segment as the weight. This reflects the greater contribution of long-distance geological segments to the horizontal conduction capacity. The expression is as follows:

[0048] In the formula, For the first The and the first Within the nth adjacent grid cell Horizontal edge weights between geological strata; The first [cell] between two adjacent grid cells The number of segments in each geological stratum; For the first The length of a geological stratification segment between two grid cells; , The first The and the first Within the first grid cell The first geological stratum The horizontal permeability of the segment.

[0049] While constructing the hydraulic conduction connectivity map, the cumulative runoff values ​​of surface depressions, valley lines, and valley lines are extracted based on digital elevation data. The principle is that surface topography is the core factor controlling the formation and confluence process of surface runoff. Surface depressions are the areas where surface water collects, valley lines are the main confluence channels for surface water, and the cumulative runoff value reflects the water collection capacity of the valley lines. The specific implementation process is as follows: A terrain-filling algorithm is used to fill local depressions in the digital elevation data, generating a depression-free digital elevation model. Local depressions can cause errors in the calculation of water flow direction; filling ensures that water can flow continuously from high-altitude areas to low-altitude areas. The water flow direction of each grid cell is calculated based on the depression-free digital elevation model, and this direction determines the destination of surface water runoff in each grid cell. The cumulative runoff value of each grid cell is calculated based on the water flow direction, indicating how many upstream grid cells' water flows will flow into the current grid cell. Grid cells with cumulative runoff values ​​exceeding a preset threshold are connected to form valley lines. A larger cumulative runoff value indicates a stronger water collection capacity at that location, making it easier to form surface runoff channels. Unfilled permanent depressions in the digital elevation data are identified and marked as surface depressions. Permanent depressions are typically long-term water accumulation areas such as lakes and reservoirs, and are important surface hydrological elements.

[0050] Finally, nodes with permeability coefficients exceeding a preset threshold in the hydraulic conduction connectivity diagram are spatially projected and matched with valley lines, assigning groundwater recharge identifiers to the matched valley segments. The principle is that underground strong permeability channels are the main pathways for groundwater to drain to the surface, while valley lines are the main interfaces for groundwater and surface water exchange. Spatial matching between the two can accurately identify the specific location and extent of groundwater recharge to surface water. Nodes with permeability coefficients exceeding the preset threshold represent areas with extremely high groundwater transport capacity; groundwater in these areas is more likely to drain to the surface through hydraulic gradients. Projecting these nodes onto the surface and matching them with valley lines determines which valley segments receive groundwater recharge, thus accurately characterizing the contribution of groundwater to surface runoff in subsequent model training. For example, in a certain watershed, nodes with a permeability coefficient exceeding 10 meters per day are identified as nodes with strong underground permeability channels. After projecting these nodes onto the surface, it is found that three valley segments overlap with the projected area. These three valley segments are then assigned groundwater recharge markers, indicating that the surface runoff in these valley segments will be continuously recharged by groundwater.

[0051] S104: Obtain rainfall distribution data and historical surface runoff observation data for the target area; map the rainfall distribution data into the initial features of the corresponding surface nodes in the hydraulic conduction connectivity graph; convert the historical surface runoff observation data into runoff ground truth labels for the observation point locations; simultaneously extract the topographic boundaries of unobserved areas as auxiliary constraints; and train the graph attention neural network; during training, in the forward propagation, force the edge weights of the strong underground infiltration channel nodes to the surface nodes corresponding to the valley lines that are given groundwater recharge identifiers to add a hydraulic offset that is positively correlated with the cumulative runoff value, so that the feature propagation is biased towards the strong surface runoff areas, thereby obtaining the runoff prediction model; Specifically, acquiring rainfall distribution data and historical surface runoff observation data for the target area is based on the principle that rainfall is the direct input source of surface runoff, and its spatial distribution directly determines the amount of runoff generated in different areas; while historical surface runoff observation data is a true record of hydrological processes, which can provide the model with a clear optimization objective. By acquiring these two types of basic data, the model is provided with input and output benchmarks, ensuring that the model can learn the causal relationship between rainfall and surface runoff.

[0052] Mapping rainfall distribution data to the initial features of corresponding surface nodes in the hydraulic conduction connectivity graph is based on the principle that there is a one-to-one correspondence between surface nodes in the hydraulic conduction connectivity graph and the spatial location of the target area. Matching discrete raster rainfall data to corresponding surface nodes according to spatial location can be converted into node features that can be processed by the graph structure. This mapping method preserves the spatial heterogeneity of rainfall and avoids the runoff estimation errors caused by the averaging of regional rainfall in traditional methods, enabling the model to accurately capture runoff differences at different locations. In addition, historical surface runoff observation data is converted into runoff ground truth labels for observation point locations. The principle is that the runoff data at observation points is a direct measurement result of the hydrological process. Using it as a label can provide a clear error basis for adjusting the model parameters. Through the ground truth labels of observation points, the model can continuously iterate and optimize, making the prediction results closer to the actual runoff values, and ensuring the prediction accuracy of the model in areas with observation data.

[0053] Simultaneously, topographic boundaries of unobserved areas are extracted as auxiliary constraints to train the graph attention neural network. The principle behind this is that surface runoff observation points are typically sparsely distributed in actual watersheds, with a large proportion of unobserved areas. Traditional models experience a significant drop in prediction accuracy in these areas. Introducing topographic physical constraints effectively addresses the problem of poor model generalization due to insufficient observation data. Specifically, identifying the areas within the target region without surface runoff observation points is crucial. Only by clearly defining the spatial extent of unobserved areas can corresponding topographic constraint features be extracted, avoiding unnecessary interference with model training in observed areas. Ridge lines, watersheds, and abrupt changes in topography are extracted as topographic boundaries in unobserved areas. Ridge lines and watersheds are natural boundaries between different catchment units, while abrupt changes in topography indicate locations where surface runoff direction changes abruptly. These topographic features effectively limit the cross-unit propagation of surface runoff. Extracting these topographic boundaries as constraints ensures that the model's feature propagation conforms to real hydrophysical laws, avoiding unrealistic cross-catchment runoff transmission. Converting topographic boundaries into constraint edges in a hydraulically connected graph leverages the fact that edges in a hydraulically connected graph represent the propagation channels of hydrological processes. Transforming topographic boundaries into specific constraint edges clearly marks the locations where propagation needs to be restricted within the graph structure. This conversion seamlessly integrates topographic physical constraints into the graph neural network structure without modifying the model's core architecture, ensuring computational efficiency. During the training of the graph attention neural network, the propagation direction of features along the constraint edges is restricted, ensuring features can only propagate from high-altitude areas to low-altitude areas. This is based on the principle that surface runoff, under the influence of gravity, can only flow from high to low altitudes—a fundamental physical law of hydrological processes. By restricting the propagation direction, the model's feature propagation is forced to conform to the gravity-driven runoff motion, avoiding predictions that violate physical principles and improving the model's physical consistency. At the same time, the propagation intensity of features on the constraint edges is limited, so that the feature propagation intensity across terrain boundaries is lower than that within the same terrain unit. The principle is that terrain boundaries hinder the cross-unit transmission of surface runoff, while runoff exchange within the same terrain unit is more frequent. By limiting the propagation intensity, the model can make more use of the observation data within the same terrain unit for prediction, reduce the interference of cross-unit data, and improve the prediction accuracy of unobserved areas.

[0054] During training, in the forward propagation, a hydraulic offset positively correlated with the runoff accumulation value is added to the edge weights of the nodes from underground strong infiltration channels to the corresponding surface nodes of the valley lines that are assigned groundwater recharge labels. This biases feature propagation towards areas with strong surface runoff. The principle is that underground strong infiltration channels are the main pathways for groundwater to drain to the surface. The larger the runoff accumulation value of a valley line, the stronger its water collection capacity and the greater the amount of groundwater recharge it receives. By increasing the hydraulic offset, the contribution of groundwater to surface runoff can be strengthened, enabling the model to accurately characterize the regulatory effect of groundwater recharge on surface runoff. Especially during the dry season, when groundwater recharge is the main source of surface runoff, this design can significantly improve the prediction accuracy during the dry season. Thus, a runoff prediction model is obtained. The principle is that through the above-mentioned multi-source data input, physical constraint introduction and groundwater recharge reinforcement training, the model can simultaneously learn the coupling law of surface runoff generation and groundwater transport process. The resulting runoff prediction model not only has high accuracy in the observed area, but also maintains good generalization ability in the unobserved area. At the same time, it can accurately reflect the impact of groundwater on surface runoff and achieve accurate runoff prediction at multiple temporal and spatial scales.

[0055] For example, in a mountainous watershed, the unobserved area accounts for 70% of the total watershed area. Three main ridgelines and two abrupt topographic change lines in this area are extracted as topographic boundaries and converted into 12 constraint edges in the hydraulic conduction connectivity graph. During training, the features are restricted to propagate only from areas above 150 meters in altitude to areas below 150 meters in altitude, and the feature propagation intensity across constraint edges is set to 30% within the same cell. At the same time, the edge weights between surface nodes and underground strong seepage channel nodes corresponding to valley lines with a runoff accumulation value exceeding 1000 are increased by a hydraulic offset equal to the runoff accumulation value of the valley line segment multiplied by 0.001. After 50 rounds of training, the model's runoff prediction accuracy in the unobserved area is improved by 28%, and its dry season prediction accuracy is improved by 35%.

[0056] S105 inputs real-time rainfall distribution data into the runoff prediction model and outputs prediction results that include the location of surface runoff at multiple temporal and spatial scales.

[0057] like Figure 2 As shown, based on the same inventive concept, this embodiment provides a multi-temporal runoff prediction system based on an AI large model, including: Data acquisition module 201 is used to acquire drilling data and digital elevation data of the target area; The geological modeling module 202 is used to construct a geological model of the target area using drilling data. The drilling data is obtained by setting multiple sampling points in the target area, drilling and core sampling at each sampling point to obtain geological stratification data and porosity data for each sampling point; based on the geological stratification data and the corresponding digital elevation data, a three-dimensional geological model of the target area is established. The connectivity graph construction module 203 is used to perform density-adaptive grid division of the three-dimensional geological model based on the sampling density of drilling data; convert each geological layer within each grid cell into a graph node containing a main permeability direction vector, which is calculated based on the effective porosity and average pore size of the geological layer; construct vertical edges between nodes of different geological layers within the same grid cell, as well as horizontal edges between adjacent nodes; the weight of any edge is obtained by weighted calculation based on the effective porosity and average pore size of different geological layers in the corresponding directions of two adjacent nodes, forming a hydraulic conduction connectivity graph of the target area; simultaneously, based on digital elevation data, surface depressions, valley lines, and the cumulative confluence values ​​of valley lines are extracted, and underground strong permeability channel nodes with permeability coefficients exceeding a preset threshold in the hydraulic conduction connectivity graph are spatially projected and matched with valley lines, assigning groundwater recharge identifiers to the matched valley line segments; The model training module 204 is used to acquire rainfall distribution data and historical surface runoff observation data of the target area; it maps the rainfall distribution data into the initial features of the corresponding surface nodes in the hydraulic conduction connectivity graph, converts the historical surface runoff observation data into runoff ground truth labels of the observation point locations, and extracts the topographic boundaries of unobserved areas as auxiliary constraints to train the graph attention neural network; during training, in the forward propagation, the edge weights from the nodes of strong underground infiltration channels to the corresponding surface nodes of the valley lines that are given groundwater recharge labels are forced to add a hydraulic offset that is positively correlated with the cumulative runoff value, so that the feature propagation is biased towards the strong surface runoff areas, thereby obtaining the runoff prediction model; The results module 205 is used to input real-time rainfall distribution data into the runoff prediction model and output prediction results that include the location of surface runoff at multiple temporal and spatial scales.

[0058] like Figure 3 As shown, based on the same inventive concept, this embodiment provides an electronic device, including: Memory 302 is used to store computer programs; Processor 301 is used to implement the method steps as described in the first aspect when executing a computer program.

[0059] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0060] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-temporal and spatial runoff prediction method based on an AI large model, characterized in that, include: Acquire drilling data and digital elevation data for the target area; A geological model of the target area is constructed using the drilling data, which is obtained by setting up multiple sampling points in the target area and drilling and coring each sampling point to obtain geological stratification data and porosity data of each sampling point. Based on the geological stratification data and the corresponding digital elevation data, a three-dimensional geological model of the target area is established; Based on the sampling density of the drilling data, the three-dimensional geological model is divided into density-adaptive grids. Each geological layer within each grid cell is converted into a graph node containing a main permeability direction vector, which is calculated based on the effective porosity and average pore size of the geological layer. Vertical edges between nodes of different geological layers within the same grid cell, as well as horizontal edges between adjacent nodes, are constructed. The weight of any edge is obtained by weighted calculation based on the effective porosity and average pore size of different geological layers in the corresponding directions of two adjacent nodes, forming a hydraulic conduction connectivity graph of the target area. Simultaneously, based on the digital elevation data, surface depressions, valley lines, and the cumulative confluence values ​​of the valley lines are extracted. The nodes of underground strong permeability channels with permeability coefficients exceeding a preset threshold in the hydraulic conduction connectivity graph are spatially projected and matched with the valley lines, and groundwater recharge identifiers are assigned to the matched valley line segments. Acquire rainfall distribution data and historical surface runoff observation data for the target area; map the rainfall distribution data to the initial features of the corresponding surface nodes in the hydraulic conduction connectivity graph; convert the historical surface runoff observation data into runoff ground truth labels for the observation point locations; simultaneously extract the topographic boundaries of unobserved areas as auxiliary constraints; and train the graph attention neural network; during training, in the forward propagation, force the edge weights from the underground strong infiltration channel nodes to the surface nodes corresponding to the valley lines that have been assigned the groundwater recharge identifier to add a hydraulic offset that is positively correlated with the cumulative runoff value, so that the feature propagation is biased towards the strong surface runoff areas, thereby obtaining the runoff prediction model; The real-time rainfall distribution data is input into the runoff prediction model, and the output is a prediction result that includes the location of surface runoff at multiple temporal and spatial scales.

2. The multi-temporal runoff prediction method based on an AI large model according to claim 1, characterized in that, The lowest point of the sampling depth for each sampling point is the lowest point of the nearest surface depression or valley line.

3. The multi-temporal runoff prediction method based on an AI large model according to claim 1, characterized in that, The establishment of a three-dimensional geological model of the target area based on the geological stratification data and the corresponding digital elevation data specifically includes: Based on the lithological type and sedimentary sequence characteristics of the geological strata, the geological strata of all the sampling points are uniformly numbered and stratigraphically correlated to establish a stratigraphic framework for the target area; Under the constraints of the stratigraphic framework, spatial interpolation is performed on the top and bottom interface elevations of the same geological stratum between adjacent sampling points to generate continuous geological strata. The geological strata are smoothed to eliminate unreasonable fluctuations caused by interpolation; the continuous geological strata are then fused with the surface model generated from the digital elevation data to form an initial three-dimensional geological model of the target area. The initial three-dimensional geological model is subjected to topological relationship checks to correct errors such as overlapping, intersecting, and missing geological layers, thereby obtaining a three-dimensional geological model of the target area.

4. The multi-temporal runoff prediction method based on an AI large model according to claim 1, characterized in that, The density-adaptive mesh generation process also includes: Calculate the average drilling sampling interval for the target area; If the distance between any two sampling points within any grid cell is less than half of the average drilling sampling interval, then the grid cell is recursively subdivided; the subdivision stops when the distance between any two sampling points within all grid cells is not less than half of the average drilling sampling interval.

5. The multi-temporal runoff prediction method based on an AI large model according to claim 1, characterized in that, The process of converting each geological layer within each grid cell into a graph node containing a principal permeation direction vector specifically includes: For each geological stratum within each grid cell, the effective porosity and average pore size of the corresponding geological stratum in the horizontal and vertical directions are obtained from the pore data; the vertical direction is the direction of gravity, and the horizontal direction is the direction perpendicular to gravity. The permeability intensity in the corresponding direction is obtained by multiplying the effective porosity and the average pore size in the corresponding direction; the main permeability direction vector of each geological layer is obtained by vector merging based on the permeability intensity in different directions. Each geological stratum within each grid cell is treated as an independent unit, and the main permeability direction vector is used as the core attribute of that unit to construct the graph node.

6. The multi-temporal runoff prediction method based on an AI large model according to claim 5, characterized in that, The weight of any edge is obtained by weighted calculation of the effective porosity and average pore size in different geological layers in the corresponding directions of two adjacent nodes; For the vertical edge, the effective porosity and average pore size of each of the two adjacent nodes in the vertical direction are obtained, and the vertical permeability of the two graph nodes is calculated respectively. The permeability of the two vertical directions is weighted by the thickness ratio of the two geological layers in the grid cell to obtain the basic weight of the vertical edge. For the horizontal edge, determine the geological stratification group between two adjacent nodes that are located in the same horizontal direction; For each group of geological strata, the effective porosity and average pore size in the horizontal direction of the two nodes are obtained respectively, and the horizontal permeability intensity of each is calculated. The average value of each group is weighted and summed by the length ratio of the geological strata in each group between the two nodes to obtain the weight of the horizontal edge.

7. The multi-temporal runoff prediction method based on an AI large model according to claim 1, characterized in that, The extraction of surface depressions, valley lines, and the cumulative confluence values ​​of the valley lines based on the digital elevation data specifically includes: A terrain-filling algorithm is used to fill local depressions in the digital elevation data to generate a depression-free digital elevation model; the water flow direction of each grid cell is calculated based on the depression-free digital elevation model; the cumulative runoff value of each grid cell is calculated according to the water flow direction; grid cells with cumulative runoff values ​​exceeding a preset threshold are connected to form the valley line; and unfilled permanent depressions in the digital elevation data are identified and marked as surface depressions.

8. The multi-temporal runoff prediction method based on an AI large model according to claim 1, characterized in that, The extraction of terrain boundaries from unobserved areas as auxiliary constraints specifically includes: Identify the area within the target region where no surface runoff observation points are set up; extract the ridgeline, watershed, and topographic abrupt change line of the unobserved area as topographic boundaries; convert the topographic boundaries into constraint edges in the hydraulic conduction connectivity graph; during the training of the graph attention neural network, restrict the propagation direction of features on the constraint edges, so that features can only propagate from high-altitude areas to low-altitude areas; at the same time, restrict the propagation intensity of features on the constraint edges, so that the feature propagation intensity across topographic boundaries is lower than the feature propagation intensity within the same topographic unit.

9. A multi-temporal runoff prediction system based on an AI large-scale model, using the multi-temporal runoff prediction method based on an AI large-scale model as described in any one of claims 1 to 8, characterized in that, include: The data acquisition module is used to acquire drilling data and digital elevation data for the target area; The geological modeling module is used to construct a geological model of the target area using the drilling data. The drilling data is obtained by setting multiple sampling points in the target area and drilling and coring each sampling point to obtain geological stratification data and porosity data. Based on the geological stratification data and the corresponding digital elevation data, a three-dimensional geological model of the target area is established; The connectivity graph construction module is used to perform density-adaptive grid division on the three-dimensional geological model based on the sampling density of the drilling data; convert each geological layer within each grid cell into a graph node containing a main permeability direction vector, which is calculated based on the effective porosity and average pore size of the geological layer; construct vertical edges between nodes of different geological layers within the same grid cell, as well as horizontal edges between adjacent nodes; the weight of any edge is obtained by weighted calculation based on the effective porosity and average pore size of different geological layers in the corresponding directions of two adjacent nodes, forming a hydraulic conduction connectivity graph of the target area; simultaneously, based on the digital elevation data, surface depressions, valley lines, and the cumulative confluence values ​​of the valley lines are extracted, and underground strong permeability channel nodes with permeability coefficients exceeding a preset threshold in the hydraulic conduction connectivity graph are spatially projected and matched with the valley lines, assigning groundwater recharge identifiers to the matched valley line segments; The model training module is used to acquire rainfall distribution data and historical surface runoff observation data of the target area; map the rainfall distribution data to the initial features of the corresponding surface nodes in the hydraulic conduction connectivity graph; convert the historical surface runoff observation data into runoff ground truth labels for the observation point locations; and extract the topographic boundaries of unobserved areas as auxiliary constraints to train the graph attention neural network. During training, in the forward propagation, the edge weights from the underground strong infiltration channel nodes to the surface nodes corresponding to the valley lines that are assigned the groundwater recharge identifier are forced to add a hydraulic offset that is positively correlated with the cumulative runoff value, so that the feature propagation is biased towards the strong surface runoff areas, thereby obtaining the runoff prediction model. The results module is used to input real-time rainfall distribution data into the runoff prediction model and output prediction results that include the location of surface runoff at multiple temporal and spatial scales.

10. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing a computer program to implement the multi-temporal runoff prediction method based on an AI large model as described in any one of claims 1 to 8.

Images