River flood control influence analysis method based on graph neural network

Abstract

The invention discloses a river flood control influence analysis method based on a graph neural network, and the method comprises the following steps: S1, collecting terrain, hydrology, weather and water flow data, and constructing a graph structure model; s2, generating node and edge features, and forming a graph model; s3, inputting the graph model to a Laplacian network, and generating graph feature representation; s4, combining real-time data updating features, and constructing a dynamic graph model; s5, executing neighborhood sampling, and constructing a local subgraph; s6, inputting an attention mechanism, and generating an attention graph representation; s7, inputting a graph generation module, simulating a flood situation, and generating an expansion graph structure; and S8, training the dynamic graph model, and outputting a flood range, a risk level and flood control measures. According to the invention, graph modeling and reasoning integration of flood control analysis is realized.

Description

Technical Field

[0001] This invention relates to the field of graph neural network and hydrological modeling technology, and in particular to a method for analyzing the impact of river flood control based on graph neural networks. Background Technology

[0002] In flood risk management in river areas, accurately analyzing flood propagation paths, risk levels, and control measures is a crucial prerequisite for ensuring navigation safety and the stable operation of riverside facilities. Traditional flood impact analysis methods mainly rely on numerical simulation techniques from hydrology, such as two-dimensional shallow water equation models, stormwater models, and topographic distribution models. These methods typically require a large number of physical parameter inputs and complex grid partitioning, resulting in high computational costs, poor real-time performance, and difficulty in dynamically responding to variable meteorological conditions and complex river system structures. Furthermore, while some methods attempt to incorporate machine learning techniques, such as support vector machines, random forests, or convolutional neural networks, for flood area identification or risk classification, they often fail to adequately characterize spatial structural features and temporal changes, making it difficult to effectively capture the nonlinear evolution patterns of water flow diffusion and resulting in insufficient generalization ability.

[0003] In recent years, graph neural networks have demonstrated significant advantages in processing non-Euclidean structured data and have been increasingly applied in fields such as transportation, meteorology, and hydrology. Some studies represent river networks or watershed structures as graph models and attempt to model the hydrological state of nodes using graph convolutional networks. However, most existing graph neural network methods only employ static graph structures, ignoring the impact of temporal changes in hydrological states on the graph structure and node states, thus lacking dynamic modeling capabilities. Furthermore, most graph neural network structures use fixed adjacency relationships and single-order propagation mechanisms, making it difficult to adapt to the uneven information propagation caused by complex hydraulic connections and spatial topographic differences. In addition, physical factors such as velocity gradients and water level differences between nodes are not effectively integrated into the graph structure modeling process, resulting in insufficient model perception of hydrological physical laws and affecting the physical consistency and reliability of predictions.

[0004] Therefore, how to provide a method for analyzing the flood control impact of river channels based on graph neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method for analyzing the impact of river flood control based on graph neural networks. This invention combines steps such as graph structure modeling, node and edge feature construction, improved Laplace smoothing network, dynamic graph update, adaptive neighborhood sampling, spatial and graph attention mechanism, graph generation simulation and reinforcement learning inference, and describes in detail the complete process from static hydrological geographic information modeling to dynamic flood expansion simulation. It has the advantages of strong structural modeling capability, sufficient physical information fusion and high dynamic response accuracy.

[0006] A method for analyzing the impact of river flood control based on graph neural networks according to an embodiment of the present invention includes the following steps:

[0007] S1. Collect topographic data, hydrological data, meteorological data and water flow monitoring data of the river area, and construct a graph structure model. The graph structure model contains nodes and edges. Nodes represent key locations in the river and edges represent the spatial relationships between nodes.

[0008] S2. Generate node features and edge features based on the graph structure model. Node features include water flow, meteorological and terrain information, and edge features represent spatial and temporal correlation attributes, forming a graph model.

[0009] S3. Input the graph model into the improved Laplacian smoothing network, perform smoothing processing on the node features, and generate graph feature representations;

[0010] S4. Update node and edge features based on graph feature representation, real-time hydrological data, and meteorological data to construct a dynamic graph model;

[0011] S5. Based on the dynamic graph model, perform adaptive neighborhood sampling, extract the node set and associated edges according to the spatiotemporal characteristics of water flow propagation, and construct a local subgraph structure;

[0012] S6. Input the local subgraph structure into the spatial attention mechanism and the graph attention mechanism, calculate the node weights and edge weights, and generate the attention graph representation;

[0013] S7. Use the attention graph representation of the input graph generation module to simulate a flood scenario and generate an extended graph structure;

[0014] S8. The dynamic graphical model is trained by combining historical hydrological data and real-time data. The spatiotemporal adaptive mechanism and reinforcement learning strategy are applied to output the flood expansion range, risk level and flood control measures.

[0015] Optionally, the improved Laplace smoothing network includes a multi-order Laplace kernel combination structure, an edge weight adjustment mechanism, and a fusion weight control module, wherein:

[0016] Multi-order Laplace kernel combinatorial structures are used to construct graph smoothing operators within different neighborhood ranges;

[0017] The edge weight adjustment mechanism calculates edge weights based on the water level difference, elevation difference, and flow velocity gradient between nodes;

[0018] The fusion weight control module determines the combined weights of Laplace kernels of each order based on the distribution of local structural features.

[0019] Optionally, the generation of the node features and edge features includes:

[0020] The water flow monitoring sequence, meteorological monitoring sequence and terrain attribute information corresponding to each node in the graph structure model are collected to construct node feature vectors. The node feature vectors contain the flow rate value, rainfall value and elevation value of the node location at each time step.

[0021] Based on the edge connection relationship in the graph structure model, the hydrological information and topographic attribute information of the nodes at both ends of the edge are extracted to construct the edge feature vector, which includes the elevation difference, flow difference and velocity gradient.

[0022] The elevation difference is the difference in elevation values ​​between two nodes, the flow difference is the difference in flow values ​​between two nodes at the same time step, and the velocity gradient is the flow difference between two nodes divided by the product of the spatial distance and the cross-sectional area of ​​the river segment.

[0023] Optionally, the smoothing process for the node features includes:

[0024] The node features in the input graph model are used to calculate the feature propagation results in different order neighborhood ranges through a multi-order Laplacian kernel combination structure.

[0025] Based on the water level difference, elevation difference, and flow velocity gradient between nodes connected by edges in the graph model, edge weights are generated to adjust the intensity distribution of feature propagation.

[0026] The fusion weight control module allocates fusion weights for each order of Laplacian kernel based on the distribution of local structural features of each node.

[0027] The propagation results of each level of features are weighted and combined according to the fusion weights to generate a smooth feature vector for each node.

[0028] Optionally, the construction of the dynamic graph model includes:

[0029] The graph feature representation is time-aligned with real-time hydrological and meteorological data, the sampling interval is standardized, and an input sequence organized by time step is generated.

[0030] At each time step, extract the flow rate, rainfall, and elevation values ​​of each node in the graph model to construct a node feature matrix;

[0031] The changes in flow, rainfall, and elevation of a node in continuous time steps are combined to form a node state difference sequence, which is used to describe the changes in the node state over time.

[0032] In the graph structure model, edges are defined as spatial association units connecting two nodes. The features of the two nodes connected by each edge at the current time step are extracted, and the flow difference, elevation difference, and velocity gradient of the edge are calculated to update the edge feature matrix.

[0033] Node state similarity is constructed to measure the degree of similarity between the states of two nodes at the same time step. The state similarity is jointly determined by the node state difference sequence and the spatial connectivity.

[0034] Edge weights are set based on the similarity of node states. The edge weight is the information propagation strength coefficient connecting two nodes. The higher the state similarity, the larger the edge weight value.

[0035] At each time step, the node feature matrix, edge feature matrix, and edge weight matrix are combined into a dynamic graph structure unit. The dynamic graph structure units at all time steps form a dynamic graph model.

[0036] Optionally, the adaptive neighborhood sampling includes:

[0037] At each time step, node features of the target node are extracted based on a dynamic graph model;

[0038] With the target node as the center, construct an adjacency subgraph. The adjacency subgraph consists of the target node, first-order adjacent nodes, and connecting edges. All nodes and edges retain node features and edge features. The edge features represent spatial association attributes and temporal association attributes.

[0039] Calculate the Euclidean distance between the target node and each adjacent node based on spatial coordinates;

[0040] Extract the node feature changes of the target node and its neighboring nodes in consecutive time steps, and construct a state change sequence;

[0041] The Euclidean distance is combined with the state change sequence to form a propagation correlation index, which is used to measure the intensity of the influence between nodes.

[0042] Adjacent nodes and connecting edges are sorted according to the propagation correlation index. Nodes and connecting edges with propagation correlation higher than the threshold are selected to construct the local subgraph structure corresponding to the target node.

[0043] All local subgraph structures corresponding to the target nodes are combined into a graph structure set, which serves as the input for the spatial attention mechanism and the graph attention mechanism.

[0044] Optionally, the processing of the spatial attention mechanism and the graph attention mechanism includes:

[0045] Based on the node and edge features in the local subgraph structure, construct the node feature matrix and the edge feature matrix;

[0046] Based on the spatial coordinates of the nodes and the terrain information, the spatial similarity matrix between the nodes is calculated. The spatial similarity matrix is ​​used to characterize the degree of correlation between the nodes in terms of geographical location and terrain structure.

[0047] A spatial attention matrix is ​​constructed based on the spatial similarity matrix. Each element of the spatial attention matrix represents the spatial domain attention distribution of a node to other nodes.

[0048] A graph attention matrix is ​​constructed based on connection relationships and edge features. Each element of the graph attention matrix represents the structural dependence strength of the target node on its neighboring nodes.

[0049] The spatial attention matrix and the graph attention matrix are fused to generate a joint attention weight matrix;

[0050] The node feature matrix is ​​weighted using a joint attention weight matrix to generate an attention graph representation.

[0051] Optionally, the processing of the graph generation module includes:

[0052] Input attention graph representation and flood scenario parameters to construct a set of flood simulation conditions. Flood scenario parameters include rainfall intensity, water level boundary and inflow distribution.

[0053] The flood simulation conditions are encoded into a scenario vector, and the scenario vector is concatenated with the attention graph representation to form a graph that generates the input tensor.

[0054] Construct a conditional graph convolutional structure, and perform graph structure expansion operations based on the node features and edge structure in the input tensor generated by the graph.

[0055] Based on the node update mechanism and connection strategy in the graph generation module, generate new nodes and new edges in the extended graph structure;

[0056] Set the number of expansion steps, perform graph generation operation at each step, and record the node states and edge connections in the current expanded graph structure;

[0057] After the expansion is completed, the node states and edge connections in the final expansion step are combined to form an expanded graph structure.

[0058] Optionally, S8 specifically includes:

[0059] Collect historical hydrological sequences, meteorological observation sequences, flood expansion range labels, risk level labels, and flood control measure labels to construct a training dataset;

[0060] Input historical hydrological and meteorological observation sequences into a dynamic graph model to generate node state sequences and graph structure sequences;

[0061] A spatiotemporal adaptive mechanism is constructed by setting a time adjustment vector and a spatial adjustment vector. The time adjustment vector represents the classification label of the time period in which the sample is located, and the spatial adjustment vector represents the relative position of the node in the river map.

[0062] Set up a reinforcement learning policy structure, define the state space, action space and reward function. The state space represents the sequence of node states and the sequence of graph structures. The action space includes flood expansion range prediction, risk level determination and flood control measure selection. The reward function is constructed based on the difference between the label and the model prediction results.

[0063] The policy network is trained using the policy gradient optimization method. The input is the state space, the output is the action distribution, the reward value corresponding to the action is calculated, and the network parameters are updated.

[0064] Set the number of training rounds, and repeatedly execute the prediction, scoring, and parameter update process until the policy network converges;

[0065] After training is completed, real-time hydrological and meteorological observation sequences are input, and the system infers and generates predictions of flood expansion range, risk level, and flood control measures.

[0066] The beneficial effects of this invention are:

[0067] This invention constructs a river channel modeling and flood control inference system based on graph neural networks. This system can fully integrate multi-source heterogeneous data such as topographic, hydrological, meteorological, and flow monitoring data to build a graph structure model with spatial correlation and temporal dynamics. Based on the graph structure, a multi-order Laplace kernel, edge weight adjustment mechanism, and fusion control module are introduced to achieve physical perception of node features and multi-scale smoothing, effectively preserving key hydrodynamic characteristics.

[0068] Meanwhile, dynamic graph modeling, adaptive neighborhood sampling, and spatial-graph attention mechanisms are employed to enhance the information representation capabilities of nodes and edges during spatiotemporal propagation. Combined with conditional graph generation structures and reinforcement learning inference mechanisms, this supports flood propagation prediction and flood control scheme recommendation based on scenario simulation. The overall approach improves the model's generalization ability and prediction accuracy under complex geographical structures and sudden hydrological scenarios. Attached Figure Description

[0069] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0070] Figure 1 This is a flowchart of a river flood control impact analysis method based on graph neural networks proposed in this invention;

[0071] Figure 2 This is a schematic diagram of the graph structure model construction for the river flood control impact analysis method based on graph neural networks proposed in this invention;

[0072] Figure 3 This is a dynamic graph model construction structure diagram for a river flood control impact analysis method based on graph neural networks proposed in this invention. Detailed Implementation

[0073] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0074] refer to Figure 1-3 A method for analyzing the flood control impact of river channels based on graph neural networks includes the following steps:

[0075] S1. Collect topographic data, hydrological data, meteorological data and water flow monitoring data of the river area, and construct a graph structure model. The graph structure model contains nodes and edges. Nodes represent key locations in the river and edges represent the spatial relationships between nodes.

[0076] S2. Generate node features and edge features based on the graph structure model. Node features include water flow, meteorological and terrain information, and edge features represent spatial and temporal correlation attributes, forming a graph model.

[0077] S3. Input the graph model into the improved Laplacian smoothing network, perform smoothing processing on the node features, and generate graph feature representations;

[0078] S4. Update node and edge features based on graph feature representation, real-time hydrological data, and meteorological data to construct a dynamic graph model;

[0079] S5. Based on the dynamic graph model, perform adaptive neighborhood sampling, extract the node set and associated edges according to the spatiotemporal characteristics of water flow propagation, and construct a local subgraph structure;

[0080] S6. Input the local subgraph structure into the spatial attention mechanism and the graph attention mechanism, calculate the node weights and edge weights, and generate the attention graph representation;

[0081] S7. Use the attention graph representation of the input graph generation module to simulate a flood scenario and generate an extended graph structure;

[0082] S8. The dynamic graphical model is trained by combining historical hydrological data and real-time data. The spatiotemporal adaptive mechanism and reinforcement learning strategy are applied to output the flood expansion range, risk level and flood control measures.

[0083] This invention constructs a graph structure model with key river locations as nodes and spatial relationships as edges, enabling a graphical representation of topographic, hydrological, meteorological, and flow information in river areas, forming a unified data representation structure. This structure supports the fusion of multi-source heterogeneous data in both spatial and temporal dimensions, enhancing the flexibility and expressiveness of the data structure. It provides fundamental support for subsequent graph neural network modeling and inference, improving the ability to express and model the propagation process of flood impacts.

[0084] In this embodiment, the improved Laplace smoothing network includes a multi-order Laplace kernel combination structure, an edge weight adjustment mechanism, and a fusion weight control module, wherein:

[0085] Multi-order Laplace kernel combinatorial structures are used to construct graph smoothing operators within different neighborhood ranges;

[0086] The edge weight adjustment mechanism calculates edge weights based on the water level difference, elevation difference, and flow velocity gradient between nodes;

[0087] The fusion weight control module determines the combined weights of Laplace kernels of each order based on the distribution of local structural features.

[0088] The improved Laplacian smoothing network constructed in this invention integrates Laplacian kernels with multi-order neighborhood ranges and introduces a hydro-topography-driven edge weight adjustment mechanism and a structure-aware fusion weight control mechanism, making the propagation of graph features more closely resemble the hydrological change trends in actual river environments. This structure can finely adjust the information flow intensity between key nodes, effectively improving the smooth representation of node features in the graph and supporting feature extraction under multi-scale structures.

[0089] In this embodiment, the generation of node features and edge features includes:

[0090] The water flow monitoring sequence, meteorological monitoring sequence and terrain attribute information corresponding to each node in the graph structure model are collected to construct node feature vectors. The node feature vectors contain the flow rate value, rainfall value and elevation value of the node location at each time step.

[0091] Based on the edge connection relationship in the graph structure model, the hydrological information and topographic attribute information of the nodes at both ends of the edge are extracted to construct the edge feature vector, which includes the elevation difference, flow difference and velocity gradient.

[0092] The elevation difference is the difference in elevation values ​​between two nodes, the flow difference is the difference in flow values ​​between two nodes at the same time step, and the velocity gradient is the flow difference between two nodes divided by the product of the spatial distance and the cross-sectional area of ​​the river segment.

[0093] This invention introduces key variables reflecting hydrological conditions, such as flow rate, rainfall, and elevation, during the node feature construction process. Based on the spatial topological relationships between nodes in the graph structure, it extracts equilateral features such as elevation difference, flow rate difference, and velocity gradient, comprehensively characterizing the spatial dynamics of water flow in the river channel. This method achieves detailed modeling of node and edge attributes, ensuring the integrity and physical meaning of the input features to the graph neural network, and providing structurally consistent foundational data for subsequent modeling.

[0094] In this embodiment, the smoothing processing of the node features includes:

[0095] The node features in the input graph model are used to calculate the feature propagation results in different order neighborhood ranges through a multi-order Laplacian kernel combination structure.

[0096] Based on the water level difference, elevation difference, and flow velocity gradient between nodes connected by edges in the graph model, edge weights are generated to adjust the intensity distribution of feature propagation.

[0097] The fusion weight control module allocates fusion weights for each order of Laplacian kernel based on the distribution of local structural features of each node.

[0098] The propagation results of each level of features are weighted and combined according to the fusion weights to generate a smooth feature vector for each node.

[0099] The feature smoothing method proposed in this invention utilizes multi-order Laplacian kernels to construct feature propagation paths at different scales within a graph model, and combines this with hydrological and topographic information to generate edge weights. Then, it controls the fusion ratio through local structural features to complete the smoothing operation on node features. This method effectively combines the topological information and physical properties of the graph structure, enhancing the directionality and selectivity of feature propagation, and improving the accuracy and stability of node representation.

[0100] In this embodiment, the construction of the dynamic graph model includes:

[0101] The graph feature representation is time-aligned with real-time hydrological and meteorological data, the sampling interval is standardized, and an input sequence organized by time step is generated.

[0102] At each time step, extract the flow rate, rainfall, and elevation values ​​of each node in the graph model to construct a node feature matrix;

[0103] The changes in flow, rainfall, and elevation of a node in continuous time steps are combined to form a node state difference sequence, which is used to describe the changes in the node state over time.

[0104] In the graph structure model, edges are defined as spatial association units connecting two nodes. The features of the two nodes connected by each edge at the current time step are extracted, and the flow difference, elevation difference, and velocity gradient of the edge are calculated to update the edge feature matrix.

[0105] Node state similarity is constructed to measure the degree of similarity between the states of two nodes at the same time step. The state similarity is jointly determined by the node state difference sequence and the spatial connectivity.

[0106] Edge weights are set based on the similarity of node states. The edge weight is the information propagation strength coefficient connecting two nodes. The higher the state similarity, the larger the edge weight value.

[0107] At each time step, the node feature matrix, edge feature matrix, and edge weight matrix are combined into a dynamic graph structure unit. The dynamic graph structure units at all time steps form a dynamic graph model.

[0108] In the process of constructing a dynamic graph model, this invention aligns graph features with real-time hydrological and meteorological data over time, constructing a dynamic matrix of node and edge features that evolves over time. Edge weights are adjusted based on node state similarity, enabling dynamic updates of the graph structure over time. This mechanism reflects the evolutionary patterns of node states and the spatial information transmission mechanism in a time series, making the model more closely resemble the propagation characteristics of actual flood processes.

[0109] In this embodiment, the adaptive neighborhood sampling includes:

[0110] At each time step, node features of the target node are extracted based on a dynamic graph model;

[0111] With the target node as the center, construct an adjacency subgraph. The adjacency subgraph consists of the target node, first-order adjacent nodes, and connecting edges. All nodes and edges retain node features and edge features. The edge features represent spatial association attributes and temporal association attributes.

[0112] Calculate the Euclidean distance between the target node and each adjacent node based on spatial coordinates;

[0113] Extract the node feature changes of the target node and its neighboring nodes in consecutive time steps, and construct a state change sequence;

[0114] The Euclidean distance is combined with the state change sequence to form a propagation correlation index, which is used to measure the intensity of the influence between nodes.

[0115] Adjacent nodes and connecting edges are sorted according to the propagation correlation index. Nodes and connecting edges with propagation correlation higher than the threshold are selected to construct the local subgraph structure corresponding to the target node.

[0116] All local subgraph structures corresponding to the target nodes are combined into a graph structure set, which serves as the input for the spatial attention mechanism and the graph attention mechanism.

[0117] This invention employs an adaptive neighborhood sampling mechanism to dynamically select key adjacent nodes and edges based on the spatiotemporal correlation index of water flow propagation, constructing a local subgraph structure with spatial and temporal dependencies. This method optimizes the sampling neighborhood according to propagation correlation, reducing redundant information input while maintaining the ability to capture key propagation paths, thus providing a structurally accurate input graph for subsequent attention mechanism processing.

[0118] In this embodiment, the processing of the spatial attention mechanism and the graph attention mechanism includes:

[0119] Based on the node and edge features in the local subgraph structure, construct the node feature matrix and the edge feature matrix;

[0120] Based on the spatial coordinates of the nodes and the terrain information, the spatial similarity matrix between the nodes is calculated. The spatial similarity matrix is ​​used to characterize the degree of correlation between the nodes in terms of geographical location and terrain structure.

[0121] A spatial attention matrix is ​​constructed based on the spatial similarity matrix. Each element of the spatial attention matrix represents the spatial domain attention distribution of a node to other nodes.

[0122] A graph attention matrix is ​​constructed based on connection relationships and edge features. Each element of the graph attention matrix represents the structural dependence strength of the target node on its neighboring nodes.

[0123] The spatial attention matrix and the graph attention matrix are fused to generate a joint attention weight matrix;

[0124] The node feature matrix is ​​weighted using a joint attention weight matrix to generate an attention graph representation.

[0125] This invention integrates spatial attention and graph attention mechanisms based on local subgraphs. It utilizes the spatial coordinates of nodes, terrain features, and the attribute information of connecting edges to construct spatial attention matrices and graph structure attention matrices, respectively. These matrices are then jointly weighted to form an attention graph representation that reflects the comprehensive dependencies between nodes. This mechanism effectively enhances the model's focus on key nodes and structures, providing key guidance for the graph generation process.

[0126] In this embodiment, the processing of the graph generation module includes:

[0127] Input attention graph representation and flood scenario parameters to construct a set of flood simulation conditions. Flood scenario parameters include rainfall intensity, water level boundary and inflow distribution.

[0128] The flood simulation conditions are encoded into a scenario vector, and the scenario vector is concatenated with the attention graph representation to form a graph that generates the input tensor.

[0129] Construct a conditional graph convolutional structure, and perform graph structure expansion operations based on the node features and edge structure in the input tensor generated by the graph.

[0130] Based on the node update mechanism and connection strategy in the graph generation module, generate new nodes and new edges in the extended graph structure;

[0131] Set the number of expansion steps, perform graph generation operation at each step, and record the node states and edge connections in the current expanded graph structure;

[0132] After the expansion is completed, the node states and edge connections in the final expansion step are combined to form an expanded graph structure.

[0133] The graph generation module designed in this invention constructs simulation conditions based on attention graph representation and flood scenario parameters, and uses a conditional graph convolution structure to expand the graph structure, simulating the propagation path and expansion boundary of flood in the graph structure, and realizing the dynamic generation of expanded nodes and connecting edges. By expanding the graph structure, the model can characterize complex water flow expansion processes, improving the fitting ability and spatial prediction accuracy of flood simulation scenarios.

[0134] In this embodiment, S8 specifically includes:

[0135] Collect historical hydrological sequences, meteorological observation sequences, flood expansion range labels, risk level labels, and flood control measure labels to construct a training dataset;

[0136] Input historical hydrological and meteorological observation sequences into a dynamic graph model to generate node state sequences and graph structure sequences;

[0137] A spatiotemporal adaptive mechanism is constructed by setting a time adjustment vector and a spatial adjustment vector. The time adjustment vector represents the classification label of the time period in which the sample is located, and the spatial adjustment vector represents the relative position of the node in the river map.

[0138] Set up a reinforcement learning policy structure, define the state space, action space and reward function. The state space represents the sequence of node states and the sequence of graph structures. The action space includes flood expansion range prediction, risk level determination and flood control measure selection. The reward function is constructed based on the difference between the label and the model prediction results.

[0139] The policy network is trained using the policy gradient optimization method. The input is the state space, the output is the action distribution, the reward value corresponding to the action is calculated, and the network parameters are updated.

[0140] Set the number of training rounds, and repeatedly execute the prediction, scoring, and parameter update process until the policy network converges;

[0141] After training is completed, real-time hydrological and meteorological observation sequences are input, and the system infers and generates predictions of flood expansion range, risk level, and flood control measures.

[0142] This invention introduces a spatiotemporal adaptive mechanism and reinforcement learning strategy during training and inference. It utilizes temporal and spatial adjustment vectors to fuse node states and graph structure sequences, constructing a state space, and integrates prediction error-driven policy network optimization into the reward function. This design supports joint modeling of multi-objective prediction tasks, enhancing the comprehensive reasoning capabilities for flood expansion range prediction, risk level identification, and flood control measure selection.

[0143] Example 1:

[0144] To verify the feasibility of this invention in practice, it was applied to a flood risk assessment task in a typical medium-to-large river network area. This area has a complex river channel structure and significant hydrological spatiotemporal variations. Traditional static modeling methods are insufficient to accurately depict the flood propagation process and its impact range, especially in situations with frequent sudden rainfall events and dense river system branches. The lack of intelligent models with dynamic reasoning capabilities can easily lead to problems such as large errors in flood propagation simulation and delayed deployment of flood control measures.

[0145] In practical deployment, a graph structure model was first constructed based on the region's basic topographic data (DEM resolution of 10m×10m), hydrological monitoring data from the past 5 years (sampling frequency of 30 minutes), meteorological observation data (covering 8 indicators including temperature, rainfall, wind speed, and humidity), and water flow monitoring data deployed at major river sections (a total of 78 monitoring nodes). Each node represents a key geographical location with observation records, including river confluences, sluice gate control points, and dam entrances; edges represent pairs of nodes with physical or hydrodynamic connections. Based on this graph structure model, the time-series hydrological characteristics, topographic elevation, and historical rainfall impact values ​​of each node were extracted, and the spatial distance, elevation difference, and historical flow velocity gradient of the edges were calculated to construct a complete initial graph model.

[0146] After inputting the graph model into an improved Laplace smoothing network, the propagation operator is adjusted based on the elevation and flow differences between nodes to achieve feature aggregation at different scales, resulting in a graph feature representation. Subsequently, real-time meteorological and hydrological data are introduced to dynamically update the node states, forming a dynamic graph model. Within the dynamic graph structure, adaptive neighborhood sampling based on propagation correlation indices is further executed. The local subgraph of each node is input into spatial attention and graph attention mechanisms to generate a joint attention graph representation, capturing the spatiotemporal influence relationships between nodes. Based on the constructed graph generation module and a set of flood scenario parameters (e.g., rainfall intensity of 50 mm per hour, initial water level of 3.2 meters, and inflow of 280 cubic meters per second), an extended graph structure is generated to predict the potential flood spread path and impact range.

[0147] In the comparative experiment, the method of the present invention was compared with traditional rule-based two-dimensional hydrodynamic models (such as simulation schemes based on HEC-RAS). Four real flood events were used as test samples, and the comparison results are shown in Table 1:

[0148] Table 1 Comparison of evaluation indicators for different methods in flood prediction tasks

[0149]

[0150] As shown in Table 1, in the three dimensions of extended range error control, risk level determination accuracy, and flood control measure recommendation coverage, the river flood control impact analysis method based on graph neural network proposed in this invention is superior to the comparative methods. In particular, in response to real-time rainfall change scenarios, its dynamic graph modeling and reinforcement learning inference structure significantly improve the simulation capability of flood impact range and decision generation efficiency.

[0151] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for analyzing flood control impact of a river channel based on a graph neural network, characterized in that, The method comprises the following steps: S1, collecting topographic data, hydrological data, meteorological data and water flow monitoring data of the river area, constructing a graph structure model, the graph structure model comprising nodes and edges, the nodes representing key positions of the river, and the edges representing spatial correlation between the nodes; S2, generating node features and edge features based on the graph structure model, the node features comprising water flow, meteorological and topographic information, and the edge features representing spatial and temporal correlation attributes, forming a graph model; S3, inputting the graph model into an improved Laplacian smoothing network to perform smoothing processing on the node features, and generating a graph feature representation; S4, updating the node features and edge features according to the graph feature representation, real-time hydrological data and meteorological data, and constructing a dynamic graph model; S5, performing adaptive neighborhood sampling based on the dynamic graph model, extracting a node set and associated edges according to the spatiotemporal characteristics of water flow propagation, and constructing a local subgraph structure; S6, inputting the local subgraph structure into a spatial attention mechanism and a graph attention mechanism to calculate node weights and edge weights, and generating an attention graph representation; S7, inputting the attention graph representation into a graph generation module to simulate a flood situation, and generating an extended graph structure; S8, training the dynamic graph model in combination with historical hydrological data and real-time data, applying a spatiotemporal adaptive mechanism and a reinforcement learning strategy, and outputting a flood expansion range, a risk level and a flood control measure.

2. The river flood control impact analysis method based on a graph neural network according to claim 1, characterized in that, The improved Laplacian smoothing network comprises a multi-order Laplacian kernel combination structure, an edge weight adjustment mechanism and a fusion weight control module, wherein: The multi-order Laplacian kernel combination structure is used to construct a graph smoothing operator in different neighborhood ranges; The edge weight adjustment mechanism calculates the edge weight according to the water level difference, the elevation difference and the flow velocity gradient between the nodes; The fusion weight control module determines the combination weight of each order of the Laplacian kernel according to the local structure feature distribution.

3. The river flood control impact analysis method based on a graph neural network according to claim 1, characterized in that, The generation of the node features and the edge features comprises: Collecting the water flow monitoring sequence, the meteorological monitoring sequence and the topographic attribute information of the position corresponding to each node in the graph structure model, constructing a node feature vector, and the node feature vector comprising the flow value, the rainfall value and the elevation value of the node position at each time step; Based on the edge connection relationship in the graph structure model, extracting the hydrological information and the topographic attribute information of the nodes at both ends of the edge, and constructing an edge feature vector, the edge feature vector comprising the elevation difference, the flow difference and the flow velocity gradient; The elevation difference is the difference between the elevation values of the two nodes, the flow difference is the difference between the flow values of the two nodes at the same time step, and the flow velocity gradient is the flow difference between the two nodes divided by the product of the spatial distance and the river section area.

4. The river flood control impact analysis method based on a graph neural network according to claim 1, characterized in that, The smoothing processing of the node features comprises: Inputting the node features in the graph model, calculating the feature propagation results in different order neighborhood ranges through the multi-order Laplacian kernel combination structure; Generating the edge weight based on the water level difference, the elevation difference and the flow velocity gradient between the nodes connected by the edge in the graph model, and using the edge weight to adjust the intensity distribution of the feature propagation; Through the fusion weight control module, distributing the fusion weight of each order of the Laplacian kernel according to the local structure feature distribution of each node; Combining the feature propagation results of each order according to the fusion weight to generate a smoothed feature vector of each node.

5. The river flood control impact analysis method based on a graph neural network according to claim 1, characterized in that, The construction of the dynamic graph model comprises: The graph feature representation is time-aligned with real-time hydrological data and meteorological data, and the sampling interval is unified to generate an input sequence organized by time steps; At each time step, the flow value, rainfall value, and elevation value of each node in the graph model are extracted to construct a node feature matrix; The changes in flow, rainfall, and elevation of the nodes at consecutive time steps are combined to form a node state difference sequence, which describes the changes in node state over time; The edges in the graph structure model are defined as spatial correlation units connecting two nodes, and the features of the two nodes connected by each edge at the current time step are extracted to calculate the flow difference, elevation difference, and flow velocity gradient of the edge, and update the edge feature matrix; A node state similarity is constructed to measure the closeness of the states of two nodes at the same time step, which is jointly determined by the node state difference sequence and the spatial connection relationship; The edge weight is set according to the node state similarity, which is the information propagation intensity coefficient of the two nodes connected by the edge. The higher the state similarity, the greater the edge weight value; At each time step, the node feature matrix, edge feature matrix, and edge weight matrix are combined into a dynamic graph structure unit, and all the dynamic graph structure units at different time steps form a dynamic graph model.

6. The river flood control impact analysis method based on a graph neural network according to claim 1, characterized in that, The adaptive neighborhood sampling includes: At each time step, the node features of the target node are extracted based on the dynamic graph model; A neighboring subgraph is constructed with the target node as the center, which consists of the target node, first-order neighboring nodes, and connecting edges. All nodes and edges retain node features and edge features, and edge features represent spatial correlation attributes and temporal correlation attributes; The Euclidean distance between the target node and each neighboring node is calculated according to the spatial coordinates; The changes in node features of the target node and neighboring nodes at consecutive time steps are extracted to construct a state change sequence; The Euclidean distance and state change sequence are combined into a propagation correlation indicator, which is used to measure the influence intensity between nodes; According to the propagation correlation indicator, the neighboring nodes and connecting edges are sorted, and the nodes and connecting edges with a propagation correlation higher than a threshold value are selected to construct a local subgraph structure corresponding to the target node; All local subgraph structures corresponding to the target nodes are combined into a graph structure set, which is used as input for the spatial attention mechanism and the graph attention mechanism.

7. The river flood control impact analysis method based on a graph neural network according to claim 1, characterized in that, The processing of the spatial attention mechanism and the graph attention mechanism includes: Based on the node features and edge features in the local subgraph structure, a node feature matrix and an edge feature matrix are constructed; According to the spatial coordinates and terrain information of the nodes, a spatial similarity matrix is calculated, which represents the correlation of nodes in geographical location and terrain structure; A spatial attention matrix is constructed based on the spatial similarity matrix, and each element in the spatial attention matrix represents the spatial domain attention distribution of a node to other nodes; A graph attention matrix is constructed based on the connection relationship and edge features, and each element in the graph attention matrix represents the structural dependence intensity of the target node to the neighboring nodes; The spatial attention matrix and the graph attention matrix are fused to generate a joint attention weight matrix; The joint attention weight matrix is used to perform weighted operation on the node feature matrix to generate an attention graph representation.

8. The river flood control impact analysis method based on a graph neural network according to claim 1, characterized in that, The processing of the graph generation module includes: The input attention map representation and the flood situation parameters are used to construct a set of flood simulation conditions, and the flood situation parameters include rainfall intensity, water level boundary, and inflow distribution; The set of flood simulation conditions is encoded into a situation vector, and the situation vector is spliced with the attention map representation to form a graph generation input tensor; A conditional graph convolution structure is constructed, and a graph structure expansion operation is performed according to the node features and edge structure in the graph generation input tensor; According to the node update mechanism and connection strategy in the graph generation module, new nodes and new edges in the expanded graph structure are generated; The number of expansion steps is set, and the node state and edge connection relationship in the current expanded graph structure are recorded at each step; After the expansion is completed, the node state and edge connection relationship in the final expansion step are combined to form the expanded graph structure.

9. The river flood control impact analysis method based on a graph neural network according to claim 1, characterized in that, The S8 specifically includes: Collecting historical hydrological sequences, meteorological observation sequences, flood expansion range labels, risk level labels, and flood control measures labels to construct a training data set; The historical hydrological sequences and meteorological observation sequences are input into the dynamic graph model to generate node state sequences and graph structure sequences; A spatio-temporal adaptive mechanism is constructed, and a time adjustment vector and a space adjustment vector are set, the time adjustment vector represents the classification label of the sample in the time period, and the space adjustment vector represents the relative position of the node in the river graph; A reinforcement learning strategy structure is set, and a state space, an action space, and a reward function are defined, the state space represents the node state sequence and the graph structure sequence, the action space includes flood expansion range prediction, risk level determination, and flood control measure selection, and the reward function is constructed based on the difference between the label and the model prediction result; A policy gradient optimization method is used to train the policy network, the state space is input, the action distribution is output, the reward value corresponding to the action is calculated, and the network parameters are updated; The number of training rounds is set, and the prediction, scoring, and parameter updating process is executed in a loop until the policy network converges; After the training is completed, the real-time hydrological sequence and meteorological observation sequence are input, and the flood expansion range, risk level, and flood control measure prediction results are inferred and generated.

Images