Intelligent runoff prediction method and system based on ai large model dynamic scene

Abstract

The application relates to the technical field of runoff prediction, and provides an intelligent runoff prediction method and system based on an AI large model dynamic scene, which comprises the following steps: deeply integrating a hydrological physical mechanism into a deep learning model, constructing a directed confluence path network according to a flow direction and an elevation difference of a basin grid, taking the network as a topological constraint of information transmission of a space-time graph neural network, and making spatial modeling strictly follow a real confluence rule. Meanwhile, corresponding seepage attenuation attributes are endowed to the grid underlying surface categories in combination, runoff calculation at a grid scale is completed in advance, and the runoff and confluence process are physically decoupled. The application captures spatial heterogeneity of a basin at a grid level fine granularity, compresses invalid fitting space of a model, improves the precision and generalization ability of runoff prediction of a complex basin, and guarantees the hydrological physical rationality of a prediction result.

Description

Technical Field

[0001] This invention relates to the field of runoff prediction technology, and more specifically, to an intelligent runoff prediction method and system based on dynamic scenarios using AI large-scale models. Background Technology

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

[0003] With the increasing demand for more refined watershed hydrological forecasting, intelligent runoff prediction methods based on graph neural networks have become a research hotspot. However, existing technologies still have significant limitations in terms of the depth of physical mechanism embedding and the accuracy of spatial modeling.

[0004] In the prior art, Chinese patent CN121503200A discloses a method and device for watershed runoff simulation by integrating graph neural networks with hydrological prior knowledge. Although this scheme constructs a graph structure through watershed gridding and incorporates hydrological prior knowledge, its graph topology is only based on spatial distance to achieve time delay weighting, and the transmission of model information does not follow the convergence direction law of real water flow. At the same time, this scheme directly inputs rainfall and watershed attribute data into the model, without paying attention to the seepage attenuation problem of underlying surface classification, which makes the embedding granularity of prior knowledge coarse and unable to accurately characterize the runoff differences of different underlying surface units and the transmission characteristics of directional convergence.

[0005] Furthermore, Chinese patent CN119940609A discloses a runoff prediction method and terminal based on spatiotemporal heterogeneous graph neural network. This scheme constructs a topology structure by using monitoring stations as graph nodes, and the spatial modeling granularity remains at the station scale, which cannot be refined to the grid unit to capture the spatial heterogeneity of the watershed underlying surface. Moreover, it directly uses the original hydrological time series data as the model input, and the runoff generation process completely depends on data-driven fitting, lacking clear physical mechanism constraints. In watershed scenarios with insufficient station density and complex and diverse underlying surface types, the prediction accuracy and generalization ability will decrease significantly.

[0006] In summary, existing technologies have not yet achieved a deep integration of grid-level directed confluence topology and underlying surface runoff physical constraints, making it difficult to simultaneously ensure the physical rationality of runoff prediction results and the accuracy of refined forecasts. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent runoff prediction method and system based on dynamic scenarios using large AI models, in order to improve the aforementioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows: Firstly, this application provides an intelligent runoff prediction method based on dynamic scenarios using a large AI model, including: S101, acquire historical time-period runoff records of the target watershed, historical time-period rainfall records of each sampling point, underlying surface classification grid map, and water flow direction of each grid; S102, construct a confluence path network between grids based on the water flow direction of each grid, and assign each grid in the underlying surface classification grid diagram to a node in the confluence path network, and assign the water flow direction between adjacent grids to a directed edge in the confluence path network, with each directed edge pointing from a grid with higher elevation to an adjacent grid with lower elevation. S103: Historical time-period rainfall records are matched to the corresponding nodes of the confluence path network according to the spatial location of the sampling points to form the time-period rainfall sequence of each node; historical time-period runoff records are marked as the historical time-period output sequence of the outlet node of the confluence path network, and the outlet node is the terminal node into which all water flows in the confluence path network finally converge. S104. For each node in the confluence path network, according to the underlying surface category of the corresponding grid in the underlying surface classification grid map, assign the corresponding surface infiltration attenuation coefficient to the node; along the direction of the directed edge of the confluence path network, combine the time-period rainfall sequence of each node with the surface infiltration attenuation coefficient of the node to generate a grid-by-grid runoff time series segment, which reflects the change of the amount of water participating in the confluence after deducting infiltration in the grid over time. S105, Construct a prediction model. The prediction model is based on a spatiotemporal graph neural network architecture; the topology of the confluence path network is used as the information transmission constraint of the model; the time-series segments of the flow generation at each node are used as time-series inputs, and the time-period output sequence of the exit node is used as outputs; the prediction model is obtained through iterative training based on training and testing sets divided from historical data; the historical data includes historical flow generation time-series segments and historical time-period output sequences. S106. Obtain the time-by-time rainfall record of the target watershed during the current period. Repeat steps S103 and S104 to obtain the current runoff time series segment. Input the current runoff time series segment into the prediction model to obtain the corresponding runoff prediction result.

[0008] Furthermore, the direction of water flow is obtained based on the height change trend of each grid in the underlying surface classification grid map; the height change is obtained by calculating the gradient change based on the elevation data of the target watershed.

[0009] Furthermore, historical time-period rainfall records are matched to corresponding nodes in the confluence path network according to the spatial location of the sampling points, forming time-period rainfall sequences for each node, specifically including: Spatial matching is performed using the inverse distance weighted interpolation method. For each node to be assigned a value in the confluence path network, a preset number of sampling points that are spatially closest to the node are selected. The rainfall of each selected sampling point in the same time period is weighted and summed to obtain the rainfall of the node in that time period. The rainfall sequence of the node in time period is constructed in chronological order.

[0010] Furthermore, based on the underlying surface category of the corresponding grid cell in the underlying surface classification raster map, a surface seepage attenuation coefficient is assigned to the node, specifically including: A mapping table between underlying surface type and surface seepage attenuation coefficient is pre-constructed. The surface seepage attenuation coefficient in the mapping table is obtained by calibrating the measured rainfall-runoff process of the experimental plot corresponding to the underlying surface type. When assigning coefficients to nodes, the corresponding surface seepage attenuation coefficient is looked up from the mapping table based on the underlying surface category of the grid corresponding to the node, and the found coefficient is assigned to the node.

[0011] Furthermore, the surface seepage attenuation coefficient assigned to this node also includes: For each node, obtain the rainfall in the previous period immediately preceding the current period; when the rainfall reaches the preset rainfall threshold, reduce the surface seepage attenuation coefficient corresponding to the node according to the preset attenuation reduction ratio, and determine the runoff volume of the node in the current period based on the reduced coefficient.

[0012] Furthermore, the time-series rainfall sequence of each node is combined with the surface seepage attenuation coefficient of that node to generate a grid-by-grid runoff time series segment, specifically including: For each node, the rainfall amount for each time period in the time-period rainfall sequence is reduced according to the surface seepage attenuation coefficient corresponding to the node, and the remaining water volume after reduction is used as the runoff volume of the grid in that time period. Arrange the flow rates of the grid at different time periods in chronological order to form the flow rate sequence segment of the grid.

[0013] Furthermore, the prediction model is obtained through iterative training based on training and testing sets derived from historical data, specifically including: Historical runoff time series segments and historical time-by-time output sequences are used to construct training and testing sample sets in chronological order. The mean square error between the predicted runoff sequence and the actual runoff sequence output by the model is used as the loss, and the model parameters are iteratively updated through the backpropagation algorithm.

[0014] Secondly, this application also provides an intelligent runoff prediction system based on dynamic scenarios using a large AI model, including: The data acquisition module is used to acquire historical time-period runoff records of the target watershed, historical time-period rainfall records of each sampling point, underlying surface classification raster map, and water flow direction of each raster. The flow map construction module is used to construct the flow path network between grids based on the water flow direction of each grid. Each grid in the underlying surface classification grid map is corresponding to a node in the flow path network, and the water flow direction between adjacent grids is corresponding to a directed edge in the flow path network. Each directed edge points from the grid with higher elevation to the adjacent grid with lower elevation. The sequence generation module is used to match historical time-period rainfall records to the corresponding nodes of the confluence path network according to the spatial location of the sampling points, forming the time-period rainfall sequence of each node; and to mark the historical time-period runoff records as the historical time-period output sequence of the outlet node of the confluence path network, where the outlet node is the terminal node into which all water flows in the confluence path network finally converge. The runoff calculation module is used to assign a surface infiltration attenuation coefficient to each node in the runoff path network according to the underlying surface category of the corresponding grid in the underlying surface classification grid map. Along the direction of the directed edge of the runoff path network, the time-period rainfall sequence of each node is combined with the surface infiltration attenuation coefficient of that node to generate a grid-by-grid runoff time series segment. The runoff time series segment reflects the change of the amount of water participating in the runoff in that grid after deducting infiltration over time. The model building module is used to build a prediction model. The prediction model is based on a spatiotemporal graph neural network architecture; the topology of the confluence path network is used as the information transmission constraint of the model; the time-series segments of the flow generation at each node are used as time-series inputs, and the time-by-time output sequence of the exit node is used as outputs; the prediction model is obtained through iterative training based on training and testing sets divided from historical data; the historical data includes historical flow generation time-series segments and historical time-by-time output sequences. The runoff prediction module is used to obtain the time-by-time rainfall records of the target watershed during the current period. It is then processed by the sequence generation module and the runoff calculation module to obtain the current runoff time series segment. The current runoff time series segment is then input into the prediction model to obtain the corresponding runoff prediction results.

[0015] 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.

[0016] Fourthly, this application also provides a readable storage medium storing a computer program that, when executed by a processor, implements the method steps of the first aspect.

[0017] The beneficial effects of this invention are as follows: This invention first constructs a directed confluence path network based on the water flow direction and elevation differences of watershed grids. Each grid corresponds to a network node, and the water flow direction between adjacent grids corresponds to a unidirectional directed edge. This ensures that the information transmission path of the model strictly follows the confluence law of real water flow, giving the spatial modeling a clear physical meaning at the topological level and solving the problem of the disconnect between the existing scheme diagram structure and the actual confluence process. On this basis, rainfall data is matched to the corresponding grid nodes according to spatial location, and corresponding surface infiltration attenuation attributes are assigned according to the underlying surface type of each grid. Before inputting into the prediction model, the runoff generation calculation at the grid scale is completed to obtain the time series of water volume participating in confluence after deducting infiltration. This achieves physical decoupling between runoff generation and confluence process, accurately characterizing the runoff generation differences of different underlying surface units and significantly compressing the invalid fitting space of the model. Ultimately, a spatiotemporal graph neural network is used as the core architecture. The information transmission rules of the model are constrained by the topological structure of the confluence path network. The runoff time series of each grid is used as input for model training. The fine granularity at the grid level can fully capture the spatial heterogeneity of the watershed underlying surface. The pre-embedding of physical mechanisms also allows the model to get rid of excessive dependence on the density of monitoring stations, effectively improving the accuracy and generalization ability of runoff prediction in complex watershed scenarios, while ensuring that the prediction results have clear hydrophysical rationality. Attached Figure Description

[0018] Figure 1 A flowchart of an intelligent runoff prediction method based on dynamic scenarios of a large AI model provided by the present invention; Figure 2 A schematic diagram of an intelligent runoff prediction system based on a dynamic scenario of a large AI 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, Confluence Map Construction Module; 203, Sequence Generation Module; 204, Runoff Calculation Module; 205, Model Construction Module; 206, Runoff Prediction 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, an intelligent runoff prediction method based on a dynamic scenario of an AI large model includes: S101, acquire historical time-period runoff records of the target watershed, historical time-period rainfall records of each sampling point, underlying surface classification raster map, and water flow direction of each raster.

[0022] Specifically, historical time-period runoff records can be collected from official hydrological monitoring stations at the outlet of the target watershed. Runoff values ​​are recorded at uniform time intervals to form a continuous time series, serving as the true output label for subsequent prediction model training and providing a benchmark for model fitting. Historical time-period rainfall records at each sampling point are collected synchronously from rainfall monitoring stations spatially distributed within the watershed, covering the spatiotemporal differences in rainfall across different areas of the watershed. This is a core input element for subsequent runoff generation calculations. The underlying surface classification raster map can be generated through the interpretation of satellite remote sensing imagery. It divides the entire watershed into uniformly sized raster units and labels the land cover type corresponding to each raster, reflecting the differences in surface infiltration capacity across different regions. The flow direction of each raster is the core topological basis for constructing the confluence path network. These four types of data together constitute a complete input set integrating spatiotemporal and physical hydrological attributes, providing data support for subsequent runoff prediction that aligns with the actual hydrological patterns of the watershed, avoiding prediction bias and insufficient generalization ability caused by a single data source.

[0023] The water flow direction is derived from the height variation trend of each grid cell in the underlying surface classification grid map. This setting follows the basic hydrological principle that water flows from high to low in nature, using the topographic relief of the grid cell as the core basis for determining the flow direction. It eliminates the need for additional manual preset flow direction rules, ensuring that the calculated water flow direction perfectly matches the actual topographic features of the watershed. This effectively improves the physical realism of the subsequent confluence path network and reduces topological errors caused by manual settings. Height variation is derived from the elevation data of the target watershed by calculating gradient changes. In practice, a publicly available digital elevation model can be used as the elevation data source for the target watershed. For each grid cell, the direction of the topographic gradient change at that grid location is obtained by comparing its elevation values ​​with those of its neighboring grid cells. A technique of traversing all grid cells in the entire watershed using a fixed neighborhood window can be employed to identify the location of the most significant drop in terrain around each grid cell, thus characterizing the height variation trend. This method can accurately depict the micro-topographic relief of the watershed through grid-level elevation calculations, ensuring the spatial accuracy of the water flow direction calculation and providing a reliable spatial benchmark for the subsequent topological construction of confluence paths.

[0024] The formula for calculating the gradient change is as follows:

[0025] In the formula, Center grid To the neighboring grid Gradient change; For grid To grid The horizontal distance from the center; , The first , The elevation value of each grid cell.

[0026] S102, based on the water flow direction of each grid, construct the confluence path network between grids, and assign each grid in the underlying surface classification grid diagram to a node in the confluence path network, and assign the water flow direction between adjacent grids to directed edges in the confluence path network, with each directed edge pointing from a grid with higher elevation to an adjacent grid with lower elevation.

[0027] Specifically, this step is based on the principle of graph theory topology modeling, which transforms the discrete distributed grid hydrological units into a network structure with associated logic. When implemented, the technique of traversing and splicing the entire grid row by row can be used. All grid units in the target watershed are traversed in row priority order. Based on the water flow direction of each grid calculated in step S101, the confluence association between units is established one by one, and finally spliced ​​to form a complete confluence path network covering the entire watershed.

[0028] In practice, each grid cell in the underlying surface classification raster map is mapped to a node in the confluence path network. This mapping is based on the principle of achieving a one-to-one mapping between spatial grids and network topology units. Specifically, each grid cell is assigned a unique spatial index code, and all basic attributes of that grid cell, such as underlying surface category, elevation value, and geographic coordinates, are loaded into the attribute set of the corresponding node, making each node an independent carrier of all hydrological attributes of the corresponding grid. Then, the flow direction between adjacent grid cells is mapped to directed edges in the confluence path network, with each directed edge pointing from a grid cell with higher elevation to an adjacent grid cell with lower elevation. This setting follows the basic hydrological principle that water flows from high to low in nature. In specific implementation, an eight-neighborhood traversal method can be used. For each central grid cell, the elevation values ​​of its eight surrounding adjacent grid cells are compared sequentially. After confirming the downstream grid cell to which the water flow is directed, a unidirectional directed edge is established between the corresponding node of the central grid cell and the corresponding node of the downstream grid cell. At the same time, neighboring cells with elevations higher than the central grid cell are excluded to avoid the occurrence of reverse flow edges. Its beneficial effect is that the direction of water transfer is strictly constrained by the directed edge, which fully conforms to the confluence pattern of the actual topography of the watershed, avoids topological errors such as cyclic confluence that do not conform to hydrological and physical logic, and improves the physical authenticity of the confluence path network and the interpretability of subsequent prediction results.

[0029] S103: Historical time-period rainfall records are matched to the corresponding nodes of the confluence path network according to the spatial location of the sampling points to form the time-period rainfall sequence of each node; historical time-period runoff records are marked as the historical time-period output sequence of the outlet node of the confluence path network, and the outlet node is the terminal node into which all water flows in the confluence path network finally converge.

[0030] Specifically, this step aligns the spatial scale and data format of multi-source hydrological data with the confluence path network. Matching sampling points to corresponding nodes in the confluence path network is based on the principle that rainfall monitoring stations within a watershed are discretely distributed, far fewer in number than the raster nodes in the confluence path network. Discrete sampling data cannot directly support node-by-node runoff calculations at the raster granularity. Therefore, spatial matching expands the point-like rainfall data into area-like raster data covering the entire watershed; this achieves spatial granularity unification of rainfall data from the station scale to the raster node scale, providing a consistent input basis for subsequent node-by-node runoff calculations. In practice, a mapping index table of geographic latitude and longitude coordinates of rainfall sampling points and node spatial index codes is pre-established. Coordinate matching quickly locates the initial node to which the sampling point belongs, and then an interpolation algorithm covers all blank nodes without monitoring stations.

[0031] Then, historical time-period runoff records are marked as historical time-period output sequences of the outlet nodes of the confluence path network. The principle is that the runoff at the outlet section of the basin is the final concentrated output of the entire basin's runoff generation and confluence process. In terms of topology, this corresponds to the terminal node in the confluence path network with zero out-degree and only inflow edges. Marking the measured runoff data as the output sequence of this node can provide a unique supervised learning label for subsequent prediction models. In specific implementation, based on the full topology traversal results of the confluence path network, all nodes without downstream pointing edges are automatically identified as candidate outlet nodes. Then, combined with the actual geographical location of the basin outlet hydrological station, the unique outlet node is matched and bound, avoiding topological misalignment and data matching deviation caused by manually specifying outlet nodes.

[0032] The process involves matching historical rainfall records for each time period to corresponding nodes in the confluence path network based on the spatial location of sampling points, forming a rainfall sequence for each node. Specifically, this includes: using a distance-inverse weighted interpolation method for spatial matching. This method is based on the principle that the spatial distribution of rainfall has a natural distance correlation; areas closer together have higher similarity in rainfall characteristics. The distance-inverse weighted method uses spatial distance as the core basis for weight allocation, aligning with the natural hydrological law of spatial attenuation of rainfall. This method achieves interpolation calculations using only geographic coordinates without relying on complex terrain correction parameters or additional underlying surface data, resulting in high computational efficiency and interpolation results that can basically reflect the spatial gradient changes of rainfall in the watershed, making it suitable for large-scale batch interpolation scenarios at the raster level. Then, for each node to be assigned a value in the confluence path network, a predetermined number of sampling points spatially closest to that node are selected. The principle behind this operation is that if too many sampling points are involved in the interpolation, irrelevant rainfall characteristics from distant stations will interfere with the local interpolation results; if too few are selected, the spatial distribution differences of rainfall cannot be fully reflected. Presetting a fixed number of neighboring stations balances interpolation accuracy and computational efficiency. Finally, using the reciprocal of the distance from the node to each selected sampling point as the weight, the rainfall at each selected sampling point in the same time period is weighted and summed to obtain the rainfall at the node in that time period. This forms the time-by-time rainfall sequence for the node. The principle behind this weighting method is that sampling points that are closer to the target node are more representative of the rainfall, and therefore should be assigned a higher weight. The weighting method using the reciprocal of the distance can intuitively reflect the negative correlation between spatial distance and rainfall influence.

[0033] The formula for calculating raster rainfall interpolation is as follows:

[0034] In the formula, For the first The node corresponding to the raster is in the _ Rainfall during the period; For distance from the first The nearest grid A set consisting of rainfall sampling points; For the first The sampling point at the th sampling point Measured rainfall during the specified time period; For grid To the The Euclidean spatial distance between rainfall sampling points.

[0035] S104. For each node in the confluence path network, based on the underlying surface category of the corresponding grid in the underlying surface classification grid map, assign the corresponding surface infiltration attenuation coefficient to the node; along the direction of the directed edge of the confluence path network, combine the time-period rainfall sequence of each node with the surface infiltration attenuation coefficient of the node to generate a grid-by-grid runoff time series segment, which reflects the change of the amount of water participating in the confluence after deducting infiltration in the grid over time.

[0036] Specifically, this step is the core runoff calculation link connecting spatial rainfall input with subsequent runoff calculation. Its principle is that different land cover types have fundamental differences in pore structure, vegetation interception capacity, and topsoil permeability, directly determining the proportion of rainfall that infiltrates the ground and cannot participate in surface runoff after falling to the surface. By quantifying abstract underlying surface attributes into calculable attenuation coefficients, the differences in runoff generation capacity at different spatial locations within the watershed can be accurately characterized. This avoids the spatial heterogeneity errors caused by traditional runoff calculations using uniform parameters across the entire region, providing input data with clear physical meaning for subsequent prediction models. In practice, an ordered traversal queue is pre-established for all nodes in the runoff path network using spatial index encoding. A serial traversal method is used to sequentially read the underlying surface category attributes attached to each node. After assigning coefficient values, the values ​​are synchronously written into the node's hydrological parameter dictionary, ensuring that the parameter configuration of all grid nodes across the entire watershed is complete and without duplication.

[0037] Specifically, based on the underlying surface category of the corresponding grid in the underlying surface classification raster map, a surface seepage attenuation coefficient is assigned to each node. This involves: firstly, pre-constructing a mapping table between underlying surface categories and surface seepage attenuation coefficients. Each surface seepage attenuation coefficient in the mapping table is calibrated using measured rainfall-runoff processes from experimental plots corresponding to the underlying surface category. This mapping table serves as the core parameter benchmark for runoff generation calculation across the entire watershed. Its principle lies in the spatial transferability of runoff generation and concentration patterns on homogeneous underlying surfaces. The coefficients, calibrated based on long-term measured data from standard hydrological experimental plots, possess solid measured support and clear physical meaning, rather than being subjective empirical assignments. This ensures the reliability and spatial consistency of runoff generation calculation results from the parameter source. In practice, the underlying surface can be classified into six typical types: forest land, grassland, cultivated land, construction land, water area, and bare rock. For each type, rainfall and runoff observation data of the same type of standard experimental plot published by the domestic hydrological authorities can be selected for three consecutive years. The benchmark seepage attenuation coefficient of the corresponding type can be obtained by back-calculating the ratio of the total rainfall of a single rainfall event to the total surface runoff. The data are then entered into a mapping table to form a standardized parameter library, which can be directly called and retrieved when assigning values ​​to subsequent nodes.

[0038] Then, when assigning node coefficients, the corresponding surface seepage attenuation coefficient is looked up from the mapping table based on the underlying surface type of the corresponding raster, and the found coefficient is assigned to the node. The principle of this lookup-based assignment method is to achieve rapid parameter matching through key-value pair mapping, eliminating the need to repeatedly calculate infiltration parameters for each raster, greatly improving the parameter configuration efficiency of massive raster nodes across the entire watershed, while ensuring that the parameter values ​​of the same type of underlying surface are completely consistent, avoiding parameter deviations and spatial disorder caused by random manual assignment.

[0039] Furthermore, assigning a surface seepage attenuation coefficient to the node also includes: for each node, obtaining the rainfall in the previous period immediately preceding the current period; when the rainfall reaches a preset rainfall threshold, reducing the surface seepage attenuation coefficient corresponding to the node according to a preset attenuation reduction ratio, and determining the runoff volume of the node in the current period using the reduced coefficient. The principle of this dynamic correction mechanism is that soil moisture content has a saturation effect. When there is sufficient rainfall in the previous period, the soil pores are gradually filled with water, and the soil infiltration capacity will decrease significantly as the moisture content increases. If a fixed baseline coefficient is still used, the infiltration volume will be overestimated and the runoff volume underestimated. Therefore, by dynamically reducing the coefficient triggered by the previous rainfall, the dynamic influence of soil dryness and wetness on infiltration capacity can be restored, fully conforming to the actual runoff physical process, thereby achieving adaptive adjustment of dynamic parameters for each period and each node.

[0040] Finally, along the directed edge direction of the runoff path network, the time-series rainfall sequence of each node is combined with the surface infiltration attenuation coefficient of that node to generate a grid-by-grid runoff time series segment. The runoff time series segment reflects the change in the amount of water participating in the runoff over time after deducting infiltration in that grid. This step is the core execution link of runoff calculation. Its principle is to complete the transformation of rainfall into runoff node by node along the runoff direction, ensuring that the spatial processing order of runoff calculation is consistent with the direction of water flow, and avoiding misalignment between spatial logic and temporal logic. After deducting the amount of water that does not participate in surface runoff, such as infiltration and vegetation interception, by the infiltration attenuation coefficient, the remaining water is the effective runoff volume. The resulting time series segment can accurately depict the evolution of the runoff process of a single grid cell over time, providing standardized temporal input at the grid granularity for subsequent spatiotemporal graph neural networks. In practice, each node is processed sequentially according to the topological order of the confluence path network from upstream to downstream. For each node's time-period rainfall sequence, the effective surface seepage attenuation coefficient corresponding to the current time period is multiplied by the time-period rainfall sequence to obtain the runoff volume of the grid for that time period. After being spliced ​​in chronological order, the runoff time series segment corresponding to that node is formed. The runoff time series segments of all nodes together constitute the runoff input dataset of the entire watershed, which can be directly input into the subsequent prediction model.

[0041] The formula for calculating the production flow rate without correction is as follows:

[0042] In the formula, When using a mapping table directly, the raster In the The amount of water produced during a given time period; When directly using the mapping table, the first Each grid cell corresponds to the surface seepage attenuation coefficient of the underlying surface type; For the first The grid node at the ... Rainfall during a given period.

[0043] The expression for the dynamic correction of the seepage attenuation coefficient is:

[0044] In the formula, For the first The grid in the first The surface seepage attenuation coefficient after time-period dynamic correction; A preset rainfall threshold is used to trigger soil saturation correction; The preset attenuation reduction ratio; For grid Rainfall in the previous period of the current period.

[0045] S105, Construct a prediction model. The prediction model is based on a spatiotemporal graph neural network architecture; the topology of the confluence path network is used as the information transmission constraint of the model; the time-series segments of the flow generation at each node are used as time-series inputs, and the time-period output sequence of the exit node is used as outputs; the prediction model is obtained by iterative training based on training and testing sets divided from historical data; the historical data includes historical flow generation time-series segments and historical time-period output sequences.

[0046] Specifically, the prediction model uses a spatiotemporal graph neural network (SPNN) architecture. The rationale behind this architecture is that SPNN possesses the dual capabilities of graph structure spatial modeling and temporal dependency capture. It can adapt to the irregular spatial distribution of watershed grids and capture the dynamic evolution of runoff processes over time. Compared to traditional single-dimensional modeling networks, this architecture can simultaneously integrate the spatial topological heterogeneity and temporal dynamics of the watershed, matching the spatiotemporal coupling characteristics of hydrological processes at the model structure level. This avoids the loss of hydrological information caused by single-dimensional modeling and enhances the model's adaptability to complex watershed scenarios. In practical implementation, a network structure with alternating layers of graph convolutional modules and temporal convolutional modules can be used to extract spatiotemporal hydrological features at different scales layer by layer, ensuring the model's feature representation capabilities.

[0047] The model uses the topology of the confluence path network as a constraint for information transmission. This setting directly uses the previously constructed confluence path network as a spatial prior constraint for the model, stipulating that the model's feature information can only be transmitted unidirectionally from upstream to downstream nodes along directed edges, rather than freely propagating across the entire graph without constraints. This strictly follows the hydrological and physical laws of unidirectional water transmission along the confluence path, embedding hydrological prior knowledge into the model structure. This eliminates the need for the model to learn spatial correlation logic from massive amounts of data in an unsupervised manner. This significantly compresses the model's parameter search space, accelerates training convergence, and ensures that the model's information transmission logic fully conforms to the real confluence process, enhancing the physical interpretability of the prediction results and avoiding abnormal predictions that violate hydrological laws. In specific implementation, the node correspondences and directed edge connections of the confluence path network are transformed into a sparse adjacency matrix, which serves as the graph structure input layer of the spatiotemporal graph neural network. This directly constrains the message transmission range of the graph convolutional layer, allowing feature transmission only between upstream and downstream nodes with directed edge connections.

[0048] The system uses runoff time-series segments from each node as time-series input and time-by-time output sequences from the outlet node as output. The principle behind this input-output configuration is that the input side uses runoff time-series segments from all grid nodes across the entire watershed, providing the model with refined runoff process data covering the entire watershed space. This fully preserves the differences in the contribution of runoff from different locations to the total runoff, avoiding the loss of detailed information caused by spatial aggregation. The output side uses only the time-by-time runoff sequence from the watershed outlet node as the prediction target, accurately meeting the core business needs of runoff prediction while avoiding increased model training difficulty and dispersion of single-target accuracy caused by multiple output targets. In specific implementation, the runoff time-series segments from each node are concatenated into a graph-structured time-series input tensor according to spatial index order. Each time step corresponds to the runoff values ​​of all nodes in the entire watershed. The output layer only retains the prediction channel corresponding to the outlet node, outputting the time-by-time runoff prediction sequence of the corresponding duration.

[0049] The prediction model is based on training and testing sets derived from historical data, obtained through iterative training. The historical data includes historical runoff time-series segments and historical time-period output sequences. Specifically, the historical runoff time-series segments and historical time-period output sequences are used to construct training and testing sample sets in chronological order. This division method strictly adheres to the temporal continuity of hydrological processes, dividing the data chronologically rather than randomly, thus avoiding inflated accuracy due to test samples containing future information, ensuring the authenticity of model performance evaluation and engineering applicability. Specifically, continuous hydrological data from the previous eight years are selected as the training sample set, and continuous data from the following two years are used as the testing sample set. Each sample set uses runoff time-series segments of 12 consecutive time periods as input features and corresponding outlet runoff sequences of the next 6 time periods as labels. Multiple sets of continuous samples are constructed by sliding time windows, covering diverse hydrological scenarios with different rainfall intensities and seasons. Then, the mean squared error between the predicted runoff sequence and the actual runoff sequence is used as the loss, and the model parameters are iteratively updated using the backpropagation algorithm. This loss function applies a stronger penalty to samples with large prediction deviations, driving the model to prioritize fitting key hydrological scenarios such as high flow rates, reducing prediction errors in extreme runoff processes, and meeting the accuracy requirements of core application scenarios such as flood control and early warning. By backpropagating the loss gradient layer by layer and iteratively updating the network weights, the model's spatiotemporal feature extraction capabilities can be automatically optimized, enabling the model to gradually learn the spatiotemporal evolution of runoff generation and confluence processes in the watershed. Specifically, an initial learning rate can be set, and an adaptive gradient descent optimizer is used for iterative training. After each training round, the prediction accuracy is verified on the test set. Training is terminated early when the accuracy on the test set no longer improves after several consecutive rounds, saving the optimal model parameters, thus balancing training sufficiency and generalization performance.

[0050] The loss function is:

[0051] In the formula, The loss function value used for model training; This represents the number of samples in the batch used for model training. For the first The sample at the th Predicted runoff values ​​for each forecast period; For the first The sample at the th Measured runoff values ​​for each forecast period; This refers to the length of the future time period output by the prediction model.

[0052] S106. Obtain the time-by-time rainfall record of the target watershed during the current period. Repeat steps S103 and S104 to obtain the current runoff time series segment. Input the current runoff time series segment into the prediction model to obtain the corresponding runoff prediction result.

[0053] In this process, steps S103 and S104 are repeated to obtain the current flow generation time sequence segment. The principle is to use the same spatial interpolation and flow generation calculation logic as the training phase to ensure that the calculation rules and data dimensions of the inference input and the training input are completely matched, thus avoiding the model accuracy decay caused by data distribution shift.

[0054] like Figure 2 As shown, based on the same inventive concept, this embodiment provides an intelligent runoff prediction system based on a dynamic scenario using a large AI model, including: Data acquisition module 201 is used to acquire historical time-period runoff records of the target watershed, historical time-period rainfall records of each sampling point, underlying surface classification grid map, and water flow direction of each grid. The flow map construction module 202 is used to construct the flow path network between grids based on the water flow direction of each grid. Each grid in the underlying surface classification grid map is corresponding to a node in the flow path network, and the water flow direction between adjacent grids is corresponding to a directed edge in the flow path network. Each directed edge points from the grid with higher elevation to the adjacent grid with lower elevation. The sequence generation module 203 is used to match historical time-period rainfall records to the corresponding nodes of the confluence path network according to the spatial location of the sampling points, forming the time-period rainfall sequence of each node; and to mark the historical time-period runoff records as the historical time-period output sequence of the outlet node of the confluence path network, where the outlet node is the terminal node into which all water flows in the confluence path network finally converge. The runoff calculation module 204 is used to assign a surface infiltration attenuation coefficient to each node in the runoff path network according to the underlying surface category of the corresponding grid in the underlying surface classification grid map; along the direction of the directed edge of the runoff path network, the time-period rainfall sequence of each node is combined with the surface infiltration attenuation coefficient of the node to generate a grid-by-grid runoff time series segment, which reflects the change of the amount of water participating in the runoff in the grid after deducting infiltration over time. The model building module 205 is used to build a prediction model. The prediction model includes a spatiotemporal graph neural network architecture; the topology of the confluence path network as the information transmission constraint of the model; the time-series segments of the flow generation at each node as the time-series input, and the time-period output sequence of the exit node as the output; the prediction model is obtained through iterative training based on training and testing sets divided from historical data; the historical data includes historical flow generation time-series segments and historical time-period output sequences. The runoff prediction module 206 is used to obtain the time-by-time rainfall records of the target watershed during the current period. It is then processed by the sequence generation module and the runoff calculation module to obtain the current runoff time series segment. The current runoff time series segment is then input into the prediction model to obtain the corresponding runoff prediction results.

[0055] 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.

[0056] Based on the same inventive concept, this embodiment provides a readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, it implements the method steps as described in the first aspect.

[0057] 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.

[0058] 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 smart runoff prediction method based on dynamic scenarios using a large AI model, characterized in that, include: S101, acquire historical time-period runoff records of the target watershed, historical time-period rainfall records of each sampling point, underlying surface classification grid map, and water flow direction of each grid; S102, based on the water flow direction of each grid, construct a confluence path network between grids, and assign each grid in the underlying surface classification grid diagram to a node in the confluence path network, and assign the water flow direction between adjacent grids to a directed edge in the confluence path network, with each directed edge pointing from a grid with a higher elevation to an adjacent grid with a lower elevation. S103, the historical time-period rainfall records are matched to the corresponding nodes of the confluence path network according to the spatial location of the sampling points, forming a time-period rainfall sequence for each node; The historical time-period runoff records are marked as historical time-period output sequences of the outlet nodes of the confluence path network, where the outlet nodes are the terminal nodes into which all water flows in the confluence path network finally converge. S104, For each node in the confluence path network, according to the underlying surface category of the corresponding grid in the underlying surface classification grid map, assign the corresponding surface seepage attenuation coefficient to the node; Along the direction of the directed edges of the confluence path network, the time-period rainfall sequence of each node is combined with the surface infiltration attenuation coefficient of that node to generate a grid-by-grid runoff time series segment. The runoff time series segment reflects the change of the amount of water participating in the confluence after deducting infiltration in that grid over time. S105, Construct a prediction model, the prediction model including a spatiotemporal graph neural network as the architecture; Information transmission constraints are modeled using the topology of the aforementioned confluence path network; The prediction model takes the time-series segments of the flow generation at each node as the time-series input and the time-period output sequence of the exit node as the output. The prediction model is obtained through iterative training based on training and testing sets divided from historical data. The historical data includes historical flow generation time-series segments and historical time-period output sequences. S106. Obtain the time-by-time rainfall record of the target watershed during the current period. Repeat steps S103 and S104 to obtain the current runoff time series segment. Input the current runoff time series segment into the prediction model to obtain the corresponding runoff prediction result.

2. The intelligent runoff prediction method based on a dynamic scenario using an AI large model according to claim 1, characterized in that, The direction of water flow is obtained based on the height change trend of each grid in the underlying surface classification grid map; the height change is obtained by calculating the gradient change based on the elevation data of the target watershed.

3. The intelligent runoff prediction method based on a dynamic scenario using a large AI model according to claim 1, characterized in that, The step of matching the historical time-period rainfall records to the corresponding nodes of the confluence path network according to the spatial location of the sampling points to form the time-period rainfall sequence of each node specifically includes: Spatial matching is performed using the inverse distance weighted interpolation method. For each node to be assigned a value in the confluence path network, a preset number of sampling points that are spatially closest to the node are selected. The rainfall of each selected sampling point in the same time period is weighted and summed using the inverse distance from the node to each selected sampling point as the weight. The rainfall of the node in that time period is obtained and the time-by-time rainfall sequence of the node is constructed in chronological order.

4. The intelligent runoff prediction method based on dynamic scenarios of AI large-scale models according to claim 1, characterized in that, The step of assigning a surface seepage attenuation coefficient to a node based on the underlying surface classification grid type of the corresponding grid grid specifically includes: A mapping table between underlying surface type and surface seepage attenuation coefficient is pre-constructed. The surface seepage attenuation coefficient in the mapping table is calibrated by the measured rainfall-runoff process of the experimental plot corresponding to the underlying surface type. When assigning coefficients to nodes, the corresponding surface seepage attenuation coefficient is found from the mapping table based on the underlying surface category of the grid corresponding to the node, and the found coefficient is assigned to the node.

5. The intelligent runoff prediction method based on a dynamic scenario using a large AI model according to claim 4, characterized in that, The surface seepage attenuation coefficient assigned to the node also includes: For each node, the rainfall in the previous period immediately preceding the current period is obtained; when the rainfall reaches a preset rainfall threshold, the surface seepage attenuation coefficient corresponding to the node is reduced according to a preset attenuation reduction ratio.

6. The intelligent runoff prediction method based on a dynamic scenario using an AI large model according to claim 1, characterized in that, The step of combining the time-period rainfall sequence of each node with the surface seepage attenuation coefficient of that node to generate a grid-by-grid runoff time series segment specifically includes: For each node, the rainfall amount in each time period of the time-by-time rainfall sequence is reduced according to the surface seepage attenuation coefficient corresponding to the node, and the remaining water volume after reduction is used as the runoff volume of the grid in that time period. The flow rate of the grid at each time period is arranged in chronological order to form a flow rate time sequence segment of the grid.

7. The intelligent runoff prediction method based on a dynamic scenario using a large AI model according to claim 1, characterized in that, The prediction model is obtained through iterative training based on training and testing sets derived from historical data, and specifically includes: The historical runoff time series segments and historical time-by-time output sequences are used to construct training and testing sample sets in chronological order; the mean square error between the predicted runoff sequence and the actual runoff sequence output by the model is used as the loss, and the model parameters are iteratively updated through the backpropagation algorithm.

8. An intelligent runoff prediction system based on dynamic scenarios using a large AI model, comprising the intelligent runoff prediction method based on dynamic scenarios using a large AI model as described in claim 1, characterized in that, include: The data acquisition module is used to acquire historical time-period runoff records of the target watershed, historical time-period rainfall records of each sampling point, underlying surface classification grid map, and water flow direction of each grid. The confluence map construction module is used to construct a confluence path network between grids based on the water flow direction of each grid. Each grid in the underlying surface classification grid map is corresponding to a node in the confluence path network, and the water flow direction between adjacent grids is corresponding to a directed edge in the confluence path network. Each directed edge points from a grid with a higher elevation to an adjacent grid with a lower elevation. The sequence generation module is used to match the historical time-period rainfall records to the corresponding nodes of the confluence path network according to the spatial location of the sampling points, so as to form the time-period rainfall sequence of each node; The historical time-period runoff records are marked as historical time-period output sequences of the outlet nodes of the confluence path network, where the outlet nodes are the terminal nodes into which all water flows in the confluence path network finally converge. The runoff calculation module is used to assign a surface seepage attenuation coefficient to each node in the runoff path network according to the underlying surface category of the corresponding grid in the underlying surface classification grid map. Along the direction of the directed edges of the confluence path network, the time-period rainfall sequence of each node is combined with the surface infiltration attenuation coefficient of that node to generate a grid-by-grid runoff time series segment. The runoff time series segment reflects the change of the amount of water participating in the confluence after deducting infiltration in that grid over time. A model building module is used to build a prediction model, which includes a spatiotemporal graph neural network architecture. Information transmission constraints are modeled using the topology of the aforementioned confluence path network; The prediction model takes the time-series segments of the flow generation at each node as the time-series input and the time-period output sequence of the exit node as the output. The prediction model is obtained through iterative training based on training and testing sets divided from historical data. The historical data includes historical flow generation time-series segments and historical time-period output sequences. The runoff prediction module is used to acquire the time-by-time rainfall records of the target watershed during the current period, process them sequentially using the sequence generation module and the runoff calculation module to obtain the current runoff time series segment, and input the current runoff time series segment into the prediction model to obtain the corresponding runoff prediction result.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the intelligent runoff prediction method based on a dynamic scenario of an AI large model as described in any one of claims 1 to 7.

10. A readable storage medium, characterized in that: The readable storage medium stores a computer program, which, when executed by a processor, implements the intelligent runoff prediction method based on a dynamic scenario of an AI large model as described in any one of claims 1 to 7.

Images