Intelligent sensing management and control method and system for disaster multi-source situation

Abstract

The invention relates to a disaster multi-source situation intelligent sensing management and control method and system. According to the method, hydrometeorological and topographic data are collected, and a standardized data set is generated through space-time alignment and anomaly cleaning; constructing a directed topological graph containing node and edge attributes based on the extracted river network topological relation; designing a neural network model, and training through a physical constraint loss function embedded in a water balance principle to obtain a flood dynamic routing prediction model; inputting real-time hydrological data into the model for graph convolution operation, and predicting water level, flow and split ratio changes of each node in a future time period; and finally, carrying out submerging simulation analysis in combination with a digital elevation model, and generating a flood control scheduling scheme and risk early warning information. The deep fusion of a physical mechanism and data driving is realized, the flood propagation rule under the river network topology constraint is effectively captured by using the graph neural network, the calculation efficiency is remarkably improved while the prediction precision is ensured, and real-time and reliable decision support is provided for flood disaster prevention and control in a complex river network region.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of disaster prevention and control, and particularly relates to a disaster multi-source situation intelligent perception management and control method and system. BACKGROUND

[0002] In the field of flood disaster prevention and control in complex river network areas, accurately predicting the dynamic evolution process of floods is the core basis for making scientific flood control decisions. In typical river network regions such as the Pearl River Delta, the middle and lower reaches of the Yangtze River, and the Yellow River Basin, the river channels are crisscrossed, and the water flow movement presents significant nonlinear characteristics. During the flood evolution process, complex hydraulic phenomena such as diversion, confluence, and even backflow are common. The flood propagation path in such regions has high uncertainty. Traditional one-dimensional or two-dimensional hydrodynamic models based on fixed parameters are difficult to accurately depict the water flow interaction in multi-branch river channels, especially under extreme rainfall conditions. Model prediction bias may lead to errors in scheduling decisions.

[0003] Current mainstream flood prediction technologies mainly fall into two categories: one is the hydrodynamic model based on physical mechanisms, which simulates the flood evolution process by solving the Saint-Venant equation set. Although this type of model has clear physical meaning, it faces the problems of high computational complexity and long time consumption when dealing with large river networks, making it difficult to meet the timeliness requirements of real-time flood response. The other is the data-driven model based on statistical learning or machine learning, which improves the computational efficiency but ignores the inherent constraints of the river network spatial topology, making it difficult to effectively capture the water distribution rules at the diversion nodes, resulting in significant error amplification in multi-branch river channel prediction.

[0004] With the intensification of global climate change, extreme flood events are occurring frequently, and the existing technical system has exposed three key defects: first, traditional methods are insufficient in representing the complex topology of river networks, making it difficult to quantify the impact of dynamic changes in diversion ratios on downstream areas; second, pure data-driven models lack physical law constraints and have weak generalization ability in extreme scenarios not covered by training data; third, the process from data collection to prediction output is fragmented, and multi-source heterogeneous data cannot form a collaborative analysis capability. These defects make it difficult for existing systems to provide timely and accurate decision support in response to sudden heavy rain and floods.

[0005] In recent years, graph neural network technology has made breakthroughs in the field of spatial relationship modeling, providing a new approach to handling river network topology dependencies. At the same time, the development of remote sensing, Internet of Things, and other technologies has made real-time acquisition of hydrological, meteorological, and topographical multi-source data possible. However, how to embed physical mechanisms into data-driven models to construct a flood dynamic routing method that can adapt to complex river network topology and meet real-time prediction requirements remains a technical challenge in the field of water conservancy disaster prevention and mitigation. SUMMARY

[0006] Therefore, it is necessary to provide a disaster multi-source situation intelligent perception management and control method and system to solve the above technical problems.

[0007] In a first aspect, the application provides a disaster multi-source situation intelligent perception management and control method, comprising:

[0008] S1, based on the hydrological monitoring sites, weather radar stations and remote sensing equipment deployed in the river network area, collecting a historical multi-source heterogeneous data set composed of water level, flow, rainfall and terrain data, performing spatio-temporal alignment and outlier cleaning processing on the historical multi-source heterogeneous data set, and generating a standardized spatio-temporal data set;

[0009] S2, based on the standardized spatio-temporal data set, extracting river channel topology relationship through hydrological analysis, and constructing a directed river network topology graph containing node attribute vectors and edge attribute vectors according to the river channel topology relationship; wherein the nodes in the directed river network topology graph correspond to river junctions, and the edges in the directed river network topology graph correspond to river sections;

[0010] S3, based on the directed river network topology graph, constructing a graph neural network model; training the graph neural network model through a physical constraint loss function to obtain a flood dynamic routing prediction model;

[0011] S4, inputting real-time hydrological and meteorological data of each node in the directed river network topology graph into the flood dynamic routing prediction model, predicting the water level, flow and flow distribution ratio of each node in the future period through graph convolution operation, and obtaining flood propagation path and dynamic distribution data of flow distribution ratio;

[0012] S5, based on the flood propagation path and dynamic distribution data of flow distribution ratio, performing inundation simulation in combination with a digital elevation model to generate a flood control scheduling scheme and risk warning information.

[0013] In a second aspect, the application also provides a disaster multi-source situation intelligent perception management and control system for implementing the method of the first aspect, which comprises:

[0014] A multi-source data acquisition and standardization module is configured to collect a historical multi-source heterogeneous data set composed of water level, flow, rainfall and terrain data based on the hydrological monitoring sites, weather radar stations and remote sensing equipment deployed in the river network area, perform spatio-temporal alignment and outlier cleaning processing on the historical multi-source heterogeneous data set, and generate a standardized spatio-temporal data set;

[0015] A river network topology modeling module is configured to extract river channel topology relationship through hydrological analysis based on the standardized spatio-temporal data set, and construct a directed river network topology graph containing node attribute vectors and edge attribute vectors according to the river channel topology relationship; wherein the nodes in the directed river network topology graph correspond to river junctions, and the edges in the directed river network topology graph correspond to river sections;

[0016] The graph neural network training module is configured to construct a graph neural network model based on the directed river network topology graph, and train the graph neural network model through a physical constraint loss function to obtain a flood dynamic routing prediction model.

[0017] The flood dynamic prediction module is configured to input real-time hydro-meteorological data of each node in the directed river network topology graph into the flood dynamic routing prediction model, and predict water levels, flow rates and diversion ratios of each node in a future period through graph convolution operation to obtain flood propagation path and dynamic distribution data of the diversion ratio.

[0018] The flood simulation and dispatching decision module is configured to perform flood simulation based on the flood propagation path and dynamic distribution data of the diversion ratio, and in combination with a digital elevation model to generate a flood control dispatching scheme and risk warning information.

[0019] In a third aspect, the present application also provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the disaster multi-source situation intelligent perception management and control method according to the first aspect when executing the computer program.

[0020] In a fourth aspect, the present application also provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the disaster multi-source situation intelligent perception management and control method according to the first aspect.

[0021] The disaster multi-source situation intelligent perception management and control method and system described above can collect hydro-meteorological and terrain data through deployment of multi-source monitoring devices, generate a standardized data set through time-space alignment and abnormality cleaning, extract river network topology relationships based on the data set to construct a directed topology graph containing node and edge attributes, design a graph neural network model, train the model through a physical constraint loss function embedded with a water balance principle to obtain a flood dynamic routing prediction model, input real-time hydrological data into the model to perform graph convolution operation, predict changes in water levels, flow rates and diversion ratios of each node in a future period, and finally perform flood simulation analysis in combination with a digital elevation model to generate a flood control dispatching scheme and risk warning information. This technical solution realizes deep fusion of physical mechanisms and data-driven methods, effectively captures flood propagation rules under river network topology constraints through a graph neural network, significantly improves calculation efficiency while ensuring prediction accuracy, and provides real-time and reliable decision support for flood disaster prevention and control in complex river network areas. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments or the related art, the drawings needed in the embodiment or the related art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0023] Figure 1 A flowchart of a disaster multi-source situation intelligent perception management and control method provided by the present application is shown in the figure;

[0024] Figure 2 A flowchart of constructing a graph neural network model in an optional embodiment of the present application is shown in the figure;

[0025] Figure 3 A structural diagram of a disaster multi-source situation intelligent perception management and control system provided by the present application is shown in the figure. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical scheme and advantages of the present application clearer and more understandable, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0027] Reference Figure 1 It shows a flowchart of a disaster multi-source situation intelligent perception management and control method provided by the present application, which includes the following steps:

[0028] S1, based on the hydrological monitoring sites, weather radar stations and remote sensing equipment deployed in the river network area, collect the historical multi-source heterogeneous data set composed of water level, flow, rainfall and terrain data, perform spatio-temporal alignment and outlier cleaning processing on the historical multi-source heterogeneous data set, and generate a standardized spatio-temporal data set.

[0029] Specifically, the hydrological monitoring sites can be distributed at key nodes of the river network, such as river junctions, tributary inlets and important flood control areas. Each site is equipped with water level meters, flow meters and other equipment for real-time monitoring of water level changes, flow velocity and flow size. For example, an ultrasonic water level meter calculates the water level height by measuring the time taken for a sound wave to travel from transmission to reception; an electromagnetic flow meter measures water flow velocity using Faraday's law of electromagnetic induction. Weather radar stations monitor rainfall by transmitting electromagnetic waves and receiving reflected waves. The basic principle is that radar waves will scatter when they encounter raindrops, and the intensity of the scattered waves is proportional to the size and number of raindrops. Remote sensing equipment includes satellite remote sensing and unmanned aerial vehicle remote sensing. Satellite remote sensing uses optical sensors and radar sensors to obtain terrain data, such as the multispectral imager of the Landsat satellite, which can provide 30-meter resolution terrain images, while synthetic aperture radar (SAR) can obtain terrain information under cloudy weather, with a resolution of 1-10 meters. Unmanned aerial vehicle remote sensing has higher flexibility and resolution, and the photogrammetry equipment carried by it can obtain centimeter-level resolution terrain data, suitable for fine monitoring in small areas.

[0030] The collected multi-source data have different formats and temporal resolutions, thus requiring spatiotemporal alignment. For example, water level data from hydrological monitoring stations might be updated every 10 minutes, while remote sensing data might be updated every two days. To enable analysis of these data within the same time frame, linear interpolation or spline interpolation methods can be used to interpolate low-frequency data onto high-frequency time series. For instance, for rainfall information in remote sensing data, interpolation algorithms can be used to estimate rainfall values ​​every 10 minutes, aligning it with the temporal resolution of the hydrological monitoring data. Outliers may be caused by sensor malfunctions, data transmission errors, or extreme natural events. For example, a water level gauge might output a constant, erroneous water level value due to float jamming. To clean up these outliers, statistical methods can be used, such as calculating the mean and standard deviation of the data series and marking data points exceeding the mean ± three times the standard deviation as outliers. Alternatively, a physical plausibility check can be performed. For example, a sudden drop of several meters in river water level within a short period (e.g., 10 minutes) is unreasonable; such data points should be considered outliers and cleaned up. In addition, cross-validation can be performed using data from neighboring sites. If the data from a certain site differs significantly from the data from other neighboring sites, then that data point may be an outlier.

[0031] After spatiotemporal alignment and outlier cleansing, the data is integrated into a standardized spatiotemporal dataset. This step allows for the definition of a unified data format and storage structure. For example, water level data can be represented as a three-dimensional array, where the dimensions represent time, spatial location, and variable type, respectively. Specifically, NetCDF (Network Common Data Format) can be used to store this data, supporting efficient storage and retrieval of multidimensional arrays. In a NetCDF file, the time dimension can be represented as a one-dimensional array containing timestamps; spatial location can be represented as a two-dimensional grid of longitude and latitude; and variable types can include water level, flow rate, rainfall, etc. To ensure data consistency and readability, a detailed data dictionary can be defined, specifying the unit, dimension, measurement method, and data precision for each variable. For example, the unit for water level is meters (m), with a precision of 0.01m; the unit for flow rate is cubic meters per second (m³ / s). 3 / s), with an accuracy of 0.1m. 3 / s. In addition, data quality control and integrity checks can be performed. Quality control can be achieved by calculating data integrity and accuracy indicators, such as a missing data rate not exceeding 5% and an outlier rate below 2%. Integrity checks ensure that the data covers the entire study area and time range, without significant gaps or time periods. If missing or incomplete data is found, these gaps can be filled through interpolation or supplementary observations.

[0032] S2, based on the standardized spatio-temporal dataset, extracting river topology relationship through hydrological analysis, and constructing a directed river network topology graph containing node attribute vectors and edge attribute vectors according to the river topology relationship; wherein the nodes in the directed river network topology graph correspond to river junctions, and the edges in the directed river network topology graph correspond to river reaches.

[0033] Specifically, the core of hydrological analysis is to identify the main stem and tributaries of the river network and their connection relationship, which can be analyzed based on digital elevation model (DEM) and hydrological monitoring data. DEM provides three-dimensional information of the terrain, and river network is identified by calculating flow direction and cumulative flow. Specifically, the flow direction can be calculated using the D8 algorithm, which assigns the flow direction of each grid to one of its eight neighbors based on the principle of the lowest point in the neighborhood. For example, in a 3x3 grid window, the flow direction of the center grid points to the one with the lowest elevation among the surrounding eight grids. The cumulative flow is calculated by adding the upstream watershed area of each grid, and a larger cumulative flow can represent the main river. Combined with the actual measurement data of the hydrological monitoring station, the extracted river topology relationship can be further verified and corrected. For example, if DEM analysis shows that a certain area is a river, but hydrological monitoring data shows that there is almost no flow in that area, then the topology relationship should be adjusted to exclude that area from the river network.

[0034] The directed river network topology graph is a graph structure, where nodes represent river junctions and edges represent river reaches. Each node and edge has an attribute vector to describe its characteristics. The node attribute vector can include location coordinates (longitude and latitude), elevation, watershed area, etc.; the edge attribute vector can include river length, slope, flow capacity, etc. In the construction process, first, the nodes and edges can be connected according to the extracted river topology relationship to form a directed graph, where the direction of the edge represents the direction of the water flow. For example, if the river flows from node A to node B, then edge A→B represents this flow direction. To ensure the accuracy and completeness of the graph, detailed calculations and records of the attributes of each node and edge can be made. The length of the river reach can be obtained by calculating the geographical distance between two nodes, such as using the Haversine formula to calculate the spherical distance between two points. The slope can be calculated by the difference in elevation between two nodes divided by the length of the river reach, for example, if the elevation of node A is 100 meters and the elevation of node B is 90 meters, and the length of the river reach is 1000 meters, then the slope is (100-90) / 1000=1%. The flow capacity can be statistically analyzed based on historical flow data, such as calculating the maximum flow value of each river reach under different frequency floods, and taking it as the attribute value of the flow capacity.

[0035] S3, based on the directed river network topology graph, constructing a graph neural network model; training the graph neural network model through a physical constraint loss function to obtain a flood dynamic routing prediction model.

[0036] Specifically, the design of the graph neural network model considers the topological structure of the river network and the hydrological physical law. The model is composed of multiple graph convolution layers, and each graph convolution layer is responsible for transmitting and updating information on the nodes and edges of the river network graph. The operation of the graph convolution layer can be represented as:

[0037]

[0038] where H (l) is the node feature matrix of the l-th layer, is the adjacency matrix with self-connection, is the corresponding degree matrix, W (l) is a trainable weight matrix, and σ is an activation function. The feature vector of each node is updated in each layer by aggregating the information of its neighbor nodes, thereby capturing the spatial dependence relationship in the river network. For example, the feature vector of node a in the (l+1)-th layer will contain the feature information of its upstream nodes b and c, which helps the model understand the propagation path of flood in the river network.

[0039] To process time series data, a time convolution layer or a long short-term memory (LSTM) unit can be introduced into the graph neural network. The time convolution layer can capture local features in the time series through a sliding window mechanism, for example, a convolution window with a size of 3 can capture data features of three consecutive time steps. The LSTM unit can handle the dependence relationship in long time series, and its forget gate, input gate and output gate structure can effectively control the flow and retention of information.

[0040] In the graph neural network (GNN), the adjacency matrix A is a key matrix for representing the connection relationship between nodes in the graph structure. For a directed river network topology graph, the adjacency matrix A is a two-dimensional matrix, whose rows and columns correspond to the nodes in the graph. If there is an edge between node i and node j, the value of matrix element A ij is 1, indicating that the two nodes are connected; otherwise, the value of A ij is 0, indicating that there is no direct connection between them. In the river network scenario, each node represents a river junction, and each edge represents a river section. For example, if node a is connected to node b through a river section, the corresponding position in a (assuming a is in the i-th row and b is in the j-th column) will be marked as 1. This helps the model understand the possible flow path and direction of water in the river network, as water will flow from one junction to another through the river section.

[0041] I is an identity matrix, which is a square matrix whose main diagonal elements are all 1 and whose non-main diagonal elements are all 0. When building the graph neural network model, the identity matrix I is added to the adjacency matrix A to form The purpose is to introduce self-connection to each node. This means that each node not only receives information from its neighbor nodes but also retains its own information during the information propagation process.

[0042] During the calculation process of the graph neural network, for example, when performing graph convolution operations, each node can aggregate information from its neighbor nodes (including itself, because self-connection is added) through This matrix allows each node to aggregate information from its neighbor nodes (including itself, because self-connection is added). This enables the model to capture the dynamic propagation relationship between the junctions in the river network, as the flow of flood in the river network is not only influenced by adjacent junctions but also related to its own flow state. For example, the water level change of a node depends not only on the upstream inflow (information from neighbor nodes), but also on factors such as the node's flood discharge capacity (its own information).

[0043] When training the graph neural network model, As a component of the graph structure, it is input into the model together with the node feature matrix. The model can learn how to update the node feature vector through to better predict the dynamic routing of flood. This graph-based training method allows the model to fully utilize the topological information of the river network, improving the accuracy of the prediction.

[0044] To incorporate physical laws into model training, a physical constraint loss function is designed. This loss function not only includes traditional prediction error terms (such as mean square error), but also includes constraint terms based on hydrological physical laws. For example, the conservation of mass constraint ensures the balance of flow at river bifurcation points. Specifically, for a bifurcation node, the input flow should be equal to the sum of the output branch flows. This constraint can be expressed as:

[0045] Q in =∑Q out,i

[0046] where Q in is the input flow, and Q out,i is the flow of the output branch.

[0047] By converting physical laws into constraint terms in the loss function, the model learns not only the data patterns but also the physical laws, thereby having better generalization ability in extreme scenarios.

[0048] Model training is an optimization process that adjusts the model's parameters by minimizing the loss function. The loss function is defined as the error between the predicted value and the true value, plus the weighted sum of the physical constraint terms. For example:

[0049] L=L pred +λL phys

[0050] where L is the total loss function, which is the final loss value considering both prediction error and physical constraints, and the goal of model training is to minimize this loss function. pred is the prediction error term, such as mean squared error (MSE), N is the number of samples, representing the total number of samples used to calculate the loss function; y i is the true value of the i-th sample, which can be the actual observed hydrological data such as water level, flow, etc. is the predicted value of the i-th sample, which is the prediction result corresponding to the true value calculated by the model based on the input data. phys is the physical constraint term, such as the error of mass conservation constraint

[0051] M is the number of bifurcation points, representing the M bifurcation points in the river network that need to satisfy the mass conservation constraint; Q in,j is the input flow at the j-th bifurcation point, which is the total flow flowing into this bifurcation point from upstream; Q out,j,i is the flow of the i-th output branch at the j-th bifurcation point, representing the flow of each branch flowing out of this bifurcation point. λ is a hyperparameter that balances the weights of the two, used to balance the relative importance of the prediction error term and the physical constraint term in the total loss function. By adjusting λ, the trade-off between data fitting and adherence to physical laws by the model can be controlled.

[0052] During training, historical data is used to supervise the learning of the model. The input data includes historical hydro-meteorological data (such as water level, flow, rainfall, etc.) of each node in the directed river network topology graph, and the output data is the water level, flow and diversion ratio at the target time point. To ensure the stability and generalization ability of the model, cross-validation can be performed, for example, using the k-fold cross-validation method. The data set is divided into k subsets, each time k-1 subsets are used for training, and the remaining one subset is used for validation, repeating k times and taking the average validation result as the performance indicator of the model. In addition, the early stopping strategy can also be used, when the loss of the validation set no longer decreases for consecutive training periods, the training is stopped to prevent overfitting.

[0053] S4, input real-time hydro-meteorological data of each node in the directed river network topology graph into the flood dynamic routing prediction model, and predict the water level, flow and diversion ratio of each node in the future period through graph convolution operation, to obtain the dynamic distribution data of flood propagation path and diversion ratio.

[0054] Specifically, real-time hydro-meteorological data includes current water levels, flow rates, and rainfall amounts at each monitoring site. These data are collected in real-time through sensor networks and transmitted to the prediction system. To ensure the accuracy of the predictions, the real-time data is pre-processed, including data cleaning and format conversion, to make it consistent with the model input requirements. Data cleaning mainly removes outliers and erroneous data, for example, if the reading of a water level sensor suddenly exceeds its historical maximum value by several times, and no similar situation occurs at adjacent sites, then this data point may be an outlier and needs to be cleaned or corrected. Format conversion converts the data into a format that the model can accept, such as organizing time series data into a tensor structure required by the model input, which may include time steps, node numbers, and feature numbers. For example, the model receives input of water level, flow rate, and rainfall data for the past 1 hour (6 10-minute intervals) to form a tensor with a shape of (6, N, 3), where N is the number of nodes.

[0055] Graph convolution operation is the core operation of graph neural network models, used to propagate and update information on the nodes and edges of the river network graph. During the prediction process, the model propagates information through the graph convolution layers layer by layer, and finally obtains the water level, flow rate, and flow distribution ratio prediction values of each node in the future time period at the output layer. Taking a simple two-layer graph convolution network as an example, its operation process can be represented as:

[0056] Z = GCNLayer(GCNLayer(X, A), A)

[0057] where GCNLayer represents a graph convolution network layer. It is a basic unit in graph neural networks (GNN) for performing convolution operations on graph-structured data. X is the input node feature matrix, A is the adjacency matrix, and Z is the output prediction result. In each layer, the feature vector of a node is updated by aggregating the features of its neighbor nodes. For example, in the first layer, the feature vector of node a will combine the feature information of its upstream nodes b and c; in the second layer, the feature of node a will further combine the features of the neighbor nodes of b and c, thereby realizing multi-hop propagation of information in the river network. The prediction results include the water level, flow rate, and flow distribution ratio of each node in the future time period. For example, the model can predict the water level change trend of each node in the future 1 hour (every 10 minutes as a time step) and the flow distribution ratio at the bifurcation point. These prediction results will provide key inputs for subsequent flood simulation and flood control decision-making.

[0058] S5, based on the dynamic distribution data of flood propagation path and flow distribution ratio, combined with the digital elevation model, to perform flood simulation and generate flood control scheduling scheme and risk warning information.

[0059] Specifically, the inundation simulation is achieved by combining the predicted flood water level data with a digital elevation model (DEM) that provides detailed information about the terrain, including elevation and surface features. The basic principle is to compare the flood water level with the terrain height to determine the extent and depth of the inundated area. Specifically, for each terrain grid point, if its elevation is lower than the predicted flood water level, the point will be inundated, and the inundation depth is the difference between the water level and the elevation. For example, if the elevation of a certain grid point is 50 meters and the predicted flood water level is 52 meters, the inundation depth of the point is 2 meters. Inundation simulation can be achieved through hydrodynamic models (such as two-dimensional flood simulation models based on Saint-Venant equations) or simplified methods (such as using water level-area relationship). Hydrodynamic models can more accurately simulate the flow process of flood, including the changes of water flow velocity, direction and water depth, but the calculation complexity is higher. Simplified methods are more suitable for large-scale rapid inundation assessment, and the basic idea is to calculate the inundated area according to the flood water level and terrain height, assuming that the water flow is uniformly distributed in the horizontal direction. For example, GIS software can be used to overlay and analyze DEM and flood water level data to quickly generate an inundation range map.

[0060] According to the results of the inundation simulation, a flood control scheduling scheme can be generated, which can include scheduling recommendations for reservoirs, dredging plans for river channels, and deployment locations of temporary flood control facilities, etc. For example, if it is predicted that a certain area will be severely inundated, it can be recommended to increase the discharge capacity of the upstream reservoir to reduce the downstream water level, or to conduct emergency dredging in the key river section to improve the flow capacity of the river channel. The development of the scheduling scheme can consider multiple factors, such as protecting key areas (such as cities, farmland), reducing economic losses, ensuring personnel safety, etc. Through optimization algorithms, the optimal scheme can be selected from different scheduling schemes to minimize the impact of flood disasters. For example, an optimization model can be established with the objective function of minimizing economic losses, and the constraint conditions include the discharge capacity of the reservoir, the flow capacity of the river channel, the availability of flood control facilities, etc. Using linear programming or genetic algorithm optimization methods, the model is solved to obtain the best scheduling scheme.

[0061] In addition to the flood control scheduling scheme, risk warning information can also be generated to timely inform relevant personnel and the public. The risk warning information includes the prediction of flooded areas, the predicted flooding time and depth, the risk level, etc. For example, if a village is predicted to be flooded within the next 2 hours with a flooding depth of 1.5 meters, a high-level flood warning is issued to the village, and residents are advised to evacuate immediately. The warning information can be issued through various channels such as mobile phone messages, radio, television, etc., so that relevant personnel can take protective measures in a timely manner. To ensure the accuracy and reliability of the warning information, a multi-model fusion method can be used to comprehensively analyze the results of different flood prediction models, improving the confidence of the prediction. For example, the prediction results of physical process models and data-driven models can be combined to take their average or weighted average as the final flood prediction result, thereby reducing the uncertainty of a single model.

[0062] The above-mentioned disaster multi-source situation intelligent perception management and control method acquires hydrological and meteorological and terrain data through deployment of multi-source monitoring equipment, generates a standardized data set through time and space alignment and abnormal cleaning; extracts river network topological relations based on this to construct a directed topological graph containing node and edge attributes; further designs a graph neural network model, trains a flood dynamic routing prediction model through a physical constraint loss function embedding the principle of water balance; inputs real-time hydrological data into the model for graph convolution operation to predict changes in water level, flow, and flow distribution ratio of each node in the future period; finally, generates a flood control scheduling scheme and risk warning information through flood simulation analysis combined with a digital elevation model. The technical solution realizes deep fusion of physical mechanisms and data-driven methods, effectively captures the flood propagation law under the topological constraints of the river network using a graph neural network, significantly improves the calculation efficiency while ensuring the prediction accuracy, and provides real-time and reliable decision support for flood disaster prevention and control in complex river network areas.

[0063] Reference Figure 2 In an optional embodiment, based on the directed river network topological graph, a graph neural network model is constructed, including the following steps:

[0064] S11, taking the original node feature matrix of the directed river network topological graph as input, and combining the adjacency matrix of the directed river network topological graph, a spatial convolution module is constructed.

[0065] The spatial convolution module is used to process the topological relations in the river network topological graph, and the output of the spatial convolution module is a spatial feature matrix of each time dimension; if there is a downstream river section connection from node a to node b, the element A ab = 1 in the adjacency matrix, otherwise A ab = 0; the original node feature matrix represents the hydrological characteristics of each node.

[0066] In the spatial convolution module, graph convolution operation based on Chebyshev polynomial approximation is performed, and the expression of the graph convolution operation is:

[0067]

[0068] wherein, denotes the spatial feature matrix at time t, X t denotes the original node feature matrix at time t; T k (·) is a k-th order Chebyshev polynomial, θ k is a trainable parameter of the k-th order, and K is a polynomial order; T k (·) is defined as: T0(x) = 1, T1(x) = x, T k (x) = 2xT k-1 (x) - T k-2 (x).

[0069] wherein, L is a Laplacian matrix, representing the spatial correlation strength between nodes in the directed river network topology graph, and the calculation formula of the Laplacian matrix is L = I - D -1 / 2 AD -1 / 2 , A is an adjacency matrix, and D is a degree matrix; the diagonal elements of the degree matrix D ii = ∑ j A ij denote the out-degree of node i in the directed river network topology graph, and I is an identity matrix; λ max is the largest eigenvalue of the Laplacian matrix.

[0070] Specifically, in the river network scenario, the feature vector of each node can cover hydrological features such as water level, flow, and rainfall, forming the original node feature matrix X t . The elements A ab of the adjacency matrix A represent the connection relationship between node a and node b. If there is a downstream river segment connecting node a to node b, then A ab = 1, otherwise 0. The spatial convolution module updates the feature vector of the node by aggregating the features of the node and its neighbor nodes, thereby capturing the spatial dependency between nodes.

[0071] The spatial convolution operation is based on Chebyshev polynomial approximation, and the expression is:

[0072]

[0073] wherein, denotes the spatial feature matrix at time t, X t is the original node feature matrix, T k (·) is a k-th order Chebyshev polynomial used for polynomial approximation, θ k is a trainable parameter of the k-th order, and K is a polynomial order. The Laplacian matrix L represents the spatial correlation strength between nodes, and the calculation formula is L = I - D -1 / 2 AD-1 / 2 where D is a degree matrix, λ max is the largest eigenvalue of the Laplacian matrix.

[0074] Specifically, the diagonal elements D ii of the degree matrix D represent the out-degree of node i. When computing the Laplacian matrix, first, a normalized adjacency matrix D -1 / 2 AD -1 / 2 is computed according to the adjacency matrix A and the degree matrix D, and then it is subtracted from the identity matrix I to obtain the Laplacian matrix. Chebyshev polynomials are used to perform polynomial approximation on the Laplacian matrix to reduce computational complexity. Among them, T0(x) = 1, T1(x) = x, T k (x) = 2xT k-1 (x) - T k-2 (x). In this way, high-order polynomials can be expressed as a linear combination of low-order polynomials. The normalized Laplacian matrix is combined with the Chebyshev polynomials to transform the input node feature matrix X t . The trainable parameters θ k learn the importance of the coefficients of polynomials of different orders, thereby capturing the features of nodes and their neighbor nodes at different spatial scales. Finally, the obtained spatial feature matrix contains the updated features of the nodes at time t, which fuse the information of the nodes themselves and their neighbor nodes, and provide a basis for flood dynamic analysis in the spatial dimension.

[0075] S12, a time convolution module is constructed with the spatial feature matrix as input, and the output of the time convolution module is a hidden state time sequence;

[0076] wherein the time convolution module adopts a gated recurrent unit structure, and the operation process of the time convolution module is as follows:

[0077]

[0078] wherein r t is a reset gate, z t is an update gate, is a candidate state, h t is a final state, h t-1 is a hidden state of the previous time step, W r , W z , and W h are trainable parameter matrices, σ is a sigmoid activation function, and ⊙ represents a Hadamard product.

[0079] Specifically, the time convolution module is constructed to process time series data and capture the time-dependent relationship of node features. Its input is the spatial feature matrix and the output is a hidden state time sequence [h1, h2,..., hT The module adopts a gated recurrent unit (GRU) structure, and the operation process is as follows:

[0080] First, the reset gate r t and the update gate z t are calculated:

[0081]

[0082] where r t controls the influence of the hidden state h t-1 at the previous moment on the candidate state , and z t controls the contribution of h t-1 and to the current hidden state h t . W r and W z are trainable parameter matrices, and sigma is a sigmoid activation function that limits the output to the range (0, 1).

[0083] Next, the candidate state

[0084]

[0085] where W h is a trainable parameter matrix, and represents the Hadamard product. The candidate state reflects the feature information at the current moment.

[0086] Finally, the hidden state h t is updated:

[0087]

[0088] where h t integrates historical information and feature information at the current moment, dynamically reflecting the real-time changes of the flood in the river network. In this way, the time convolution module can effectively capture the dynamic characteristics in the time series, add a time dimension depth to the spatial feature matrix, and enable the model to understand the trend of the flood at different time points.

[0089] S13, based on the spatial convolution module and the time convolution module, an initial neural network is constructed, and a three-dimensional spatio-temporal fusion feature tensor is obtained by stacking the hidden state time series along the time dimension.

[0090] Specifically, the graph neural network framework aims to integrate spatial and temporal features, comprehensively capturing the spatiotemporal dependencies of nodes in the river network topology graph, and providing comprehensive information for flood dynamic routing prediction. Specifically, the spatial convolution module extracts the spatial dependencies between nodes through graph convolution operations to obtain a spatial feature matrix. The temporal convolution module captures the dynamic changes in the time series through the GRU structure to obtain a hidden state time series. The hidden state time series is stacked along the time dimension to form a three-dimensional spatiotemporal fusion feature tensor, with dimensions of T x N x F (T is the number of time steps, N is the number of nodes, and F is the feature dimension). This feature tensor integrates the spatial and temporal features of nodes at different time steps, providing a comprehensive perspective for flood dynamic analysis in the spatial and temporal dimensions.

[0091] For example, when predicting flood propagation paths and diversion ratios, the three-dimensional spatiotemporal fusion feature tensor enables the model to consider both the historical water level changes of nodes (temporal features) and their positions and connection relationships in the river network (spatial features), thereby more accurately predicting the dynamic evolution of floods. In this way, not only is the prediction accuracy improved, but the model's ability to understand and describe the complex dynamic behavior of floods is also enhanced.

[0092] The entire process of constructing the graph neural network framework organically integrates the outputs of the spatial convolution module and the temporal convolution module to form a model that can simultaneously process spatial and temporal data, providing a comprehensive and powerful tool for flood dynamic analysis in the spatial and temporal dimensions, and laying the foundation for subsequent flood dynamic routing prediction.

[0093] In this embodiment, flood dynamic routing prediction is achieved by constructing a graph neural network model, and the technical effects mainly lie in the ability to accurately simulate flood propagation paths and flow distribution in complex river networks. This method innovatively integrates spatial convolution modules and temporal convolution modules, enabling the model to simultaneously process spatial topology structures and time series data. The spatial convolution module, with the help of graph convolution operations, can fully capture the complex spatial relationships and water flow propagation directions between nodes in the river network, while the temporal convolution module uses gated recurrent units to effectively grasp the dynamic change rules of flood flow and other hydrological features over time. Compared with traditional methods, this technology greatly improves the accuracy of flood prediction, providing more reliable basis for aerial flood control and risk assessment, thereby playing a key role in flood disaster prevention and control in complex river network regions and effectively assisting relevant departments in formulating accurate flood control strategies in advance to reduce losses caused by floods.

[0094] In an alternative embodiment, the graph neural network model is trained by a physical constraint loss function to obtain a flood dynamic routing prediction model, including the following steps:

[0095] S21, a continuity equation constraint term is constructed based on the principle of water balance, and the expression of the continuity equation constraint term is:

[0096]

[0097] where M1 represents the upstream node set of node n in the directed river network topology graph, M2 represents the downstream node set of node n in the directed river network topology graph, Q t (i, n) represents the predicted flow value from node i to node n at time t, Q t (n, j) represents the predicted flow from node n to node j at time t, A t (n) represents the predicted water cross-sectional area at node n at time t, Δt represents the time step, Δx n represents the average length of all river sections connected to node n, q t (n) is the predicted lateral inflow of the river section where node n is located, N represents the total number of nodes in the directed river network topology graph, and T represents the number of steps of the total prediction period.

[0098] Specifically, the core of this expression is based on the principle of water balance, which ensures that the flow input and output at each node remains balanced. Specifically, the left side of the equation represents the rate of change of water volume at node n, and the right side represents the difference between the inflow and outflow of the node and the lateral inflow. In this way, the model can capture the dynamic changes of flood in the river network and ensure that the prediction results comply with the physical laws.

[0099] In practical applications, A t (n) can be calculated by the water level and the geometric shape of the water cross section. For example, assuming that the water cross section of the river section is rectangular, the water cross section area A t (n) can be represented as the product of the water level height h t (n) and the river section width b(n), that is, A t (n) = h t (n)·b(n). The flow Q t (i, n) and Q t (n, j) can be measured by flow sensors or calculated based on flow velocity and water cross section area. The lateral inflow q t (n) can include rainfall runoff, groundwater recharge, etc.

[0100] S22, combine the prediction error term of the initial neural network with the continuity equation constraint term to form a comprehensive loss function; the expression of the comprehensive loss function is:

[0101]

[0102] where α and β are adjustment coefficients, y t,n is the measured flow value of node n at time t, and is the predicted flow value of node n at time t by the initial neural network, and N is the total number of nodes.

[0103] Specifically, the prediction error term is the mean squared error (MSE), which measures the difference between the model's predicted value and the actual value. The continuity equation constraint term L cont ensures that the model's prediction results comply with the water balance principle. By adjusting the coefficients α and β, the weights of the two terms can be balanced during the training process, so that the model can fit the data while following the physical laws.

[0104] In practical applications, the selection of adjustment coefficients α and β can be adjusted according to specific circumstances. The optimal α and β values can be determined by cross-validation method. For example, the data set can be divided into training set and validation set, different α and β combinations are tried, and the combination that makes the validation set loss minimum is selected as the final adjustment coefficient.

[0105] S23, extracting a historical flood event data set from the standardized spatio-temporal data set; based on the historical flood event data set, training the initial neural network combined with the comprehensive loss function to obtain a flood dynamic routing prediction model.

[0106] Specifically, the historical flood event data set contains detailed data of multiple flood events, including water level, flow, rainfall, etc. These data are preprocessed, including missing value filling, outlier processing and spatio-temporal alignment, to ensure the quality and consistency of the data.

[0107] During the training process, the parameters of the initial neural network (such as the trainable parameters in the spatial convolution module and the temporal convolution module) are updated through the optimization of the comprehensive loss function L total Optimization algorithms include stochastic gradient descent (SGD) and Adam optimizer, etc. The goal of training is to minimize the comprehensive loss function, so that the model can accurately fit the historical data when predicting the flow, and also meet the physical constraints of water balance.

[0108] For example, assuming that the Adam optimizer is used for training with a learning rate of 0.001 and a batch size of 32. The training process can be divided into multiple epochs, and each epoch traverses the entire training set once. In each epoch, a batch of data of the same size is randomly selected for forward propagation, the comprehensive loss function is calculated, and then the model parameters are updated through backpropagation. The training process can continue until the validation set loss no longer decreases, to avoid overfitting.

[0109] In this way, the final flood dynamic routing prediction model can effectively integrate multi-source heterogeneous data, use advanced graph neural network technology and physical constraints, realize dynamic perception and intelligent control of flood disasters, and provide scientific basis and technical support for flood control and disaster reduction.

[0110] In this embodiment, the reliability and physical rationality of the flood prediction model are significantly improved by introducing a physical constraint loss function to train the graph neural network model. The innovation lies in incorporating physical laws such as water balance principles into the model training process in the form of continuity equation constraint terms, ensuring that the model learns historical data patterns while strictly following the physical laws of flood flow. By combining with the prediction error term of the initial neural network to form a comprehensive loss function, the model can automatically balance the relationship between data fitting and physical law compliance during training. This technology effectively avoids the problem of over-reliance on data and neglect of physical reality, improves the model's generalization ability and long-term prediction stability for flood dynamics, making the prediction results more reliable under various complex scenarios, and providing more solid scientific support for flood control decisions.

[0111] In an alternative embodiment, real-time hydro-meteorological data of each node in the directed river network topology graph is input into the flood dynamic routing prediction model, and the water level, flow and diversion ratio of each node in the future period are predicted through graph convolution operation to obtain dynamic distribution data of flood propagation path and diversion ratio, including the following steps:

[0112] S31, based on the real-time water level monitoring data and real-time flow monitoring data of each node in the directed river network topology graph, and combining the real-time rainfall distribution data obtained by the weather radar, the format standardization processing is performed to generate the real-time input feature matrix.

[0113] Specifically, the real-time water level monitoring data and real-time flow monitoring data come from hydrological monitoring sites deployed in the river network area. These sites are distributed at key locations of the river, such as junctions, tributary entrances and important flood control areas. The real-time rainfall distribution data obtained by the weather radar provides the rainfall intensity and distribution in the basin. These data have different formats and time resolutions, so format standardization processing is performed.

[0114] Format standardization processing includes the following aspects:

[0115] 1) Time alignment: unify different time resolution data to the same time interval. For example, water level and flow data may be updated every 10 minutes, while rainfall data may be updated every 5 minutes. By interpolation method (such as linear interpolation or spline interpolation), low-frequency data is interpolated to high-frequency time series to ensure that all data have corresponding values at the same time point.

[0116] 2) Unit uniformity: ensure that the units of all data are consistent. For example, the unit of water level data is unified to meters (m), the unit of flow data is unified to cubic meters per second (m 3 / s), and the unit of rainfall data is unified to millimeters per hour (mm / h).

[0117] 3) Missing value treatment: Fill in the missing parts of the data. Methods that can be used include interpolation using data from adjacent time points, statistical filling using historical data from the same site, or estimation based on data from surrounding sites.

[0118] 4) Outlier detection and correction: Identify and handle outliers in the data. Outliers can be caused by sensor failure, data transmission errors, or extreme natural events. Outliers are identified by statistical methods (such as calculating the mean and standard deviation of the data sequence, and marking data points that are more than ±3 times the standard deviation from the mean as outliers) or reasonable checks based on physical laws (such as a sudden drop in water level by several meters in a short time is unreasonable), and are corrected or removed.

[0119] After the above processing, the real-time input feature matrix is generated. Each row of the matrix represents a node, each column represents a feature (such as water level, flow, rainfall, etc.), and the elements in the matrix are the real-time monitoring data values after standardization processing. For example, assuming there are 5 nodes, each node has 3 features (water level, flow, rainfall), the dimension of the generated real-time input feature matrix is 5x3. This matrix provides the current hydro-meteorological state of each node in the river network topology for the flood dynamic routing prediction model, and is the basis for future flood prediction.

[0120] S32, input the real-time input feature matrix into the spatial convolution module of the flood dynamic routing prediction model, and propagate the hydrological state information on the river network topology through graph convolution operation to generate a node spatial feature matrix.

[0121] Specifically, the spatial convolution module uses the graph convolution layer in the graph neural network to process the topological relationship of the river network and capture the spatial dependency between nodes. The core of the graph convolution operation is to aggregate the feature information of the node and its neighbor nodes to update the feature vector of the node. In the river network scenario, the features representing each node (such as river junctions) depend not only on its own real-time hydrological state, but also on the influence of its upstream and downstream neighbor nodes.

[0122] The specific process of graph convolution operation is as follows:

[0123] 1) Initialization: The real-time input feature matrix is taken as the initial input of the graph convolution layer, and the feature vector of each node contains its own water level, flow, rainfall, etc.

[0124] 2) Feature aggregation: For each node, collect the feature information of its neighbor nodes (determined according to the directed river network topology graph). Neighbor nodes include nodes in the upstream water direction and downstream water direction, because the propagation of flood in the river has a clear directionality.

[0125] 3) Weight learning: By training the graph neural network model, the weights of node features and neighbor node features are learned. These weights determine the importance of different node features in the aggregation process. For example, the traffic features of upstream nodes may have a greater impact on the future traffic prediction of the current node, so they can be assigned a higher weight.

[0126] 4) Feature update: The features of the node itself and the features of the neighbor nodes are weighted and summed according to the learned weights to generate a new feature vector. This process can be seen as the propagation of hydrological state information on the river network topology, simulating the dynamic propagation process of flood in the river network.

[0127] 5) Nonlinear transformation: To increase the expression ability of the model, a nonlinear activation function (such as ReLU) is applied after feature update. This helps the model capture complex nonlinear relationships in the hydrological process, such as the relationship between flow and water level, which may be influenced by factors such as river terrain, roughness, etc.

[0128] After the graph convolution operation, a node spatial feature matrix is generated. This matrix still maintains the feature representation of each node, but the feature vector has integrated information from neighbor nodes, more comprehensively reflecting the spatial distribution and propagation trend of flood in the river network. For example, the new feature vector of a node may not only contain its own water level and flow information, but also integrate the flow increase trend of upstream nodes and the water level change of downstream nodes, providing more rich information for flood dynamic analysis in the spatial dimension.

[0129] S33, input the node spatial feature matrix into the time convolution module of the flood dynamic routing prediction model, capture the evolution law of hydrological time series through the gated recurrent unit, and generate a spatio-temporal fusion feature matrix.

[0130] Specifically, the time convolution module adopts a gated recurrent unit (GRU) structure, which aims to process time series data and capture the dynamic changes of node features over time. Its core function is to combine the feature propagation results in the spatial dimension and further integrate the information in the time dimension, so as to more accurately simulate the dynamic evolution process of flood.

[0131] The time convolution module receives the node spatial feature matrix from the spatial convolution module as input. This module uses the GRU structure to process time series, and the specific process is as follows:

[0132] 1) Initialize the hidden state: At the starting time of the time series, initialize the hidden state vector. The hidden state vector is used to store historical information in the time series, and the initial value can be set to all zeros.

[0133] 2) Time step loop: For each time step in the time series, perform the following operations:

[0134] a. Input gating: combines the input features of the current time step (node spatial features from the spatial convolution module) and the hidden state of the previous time step to calculate the values of the reset gate and the update gate. The reset gate determines the degree of influence of the previous time hidden state on the candidate state; the update gate determines the contribution proportion of the previous time hidden state and the candidate state to the current hidden state.

[0135] b. Candidate state calculation: based on the value of the reset gate and the hidden state of the previous time, the candidate state is calculated. The candidate state reflects the feature information at the current time, which combines the influence of the current input and historical information.

[0136] c. Hidden state update: according to the value of the update gate, the previous time hidden state and the candidate state at the current time are fused to update the hidden state vector. The hidden state vector captures the dynamic characteristics in the time series, and is constantly updated to reflect the latest time evolution law as the time step advances.

[0137] d. Spatio-temporal feature fusion: after processing through multiple time steps, the hidden state vector contains evolution information in the time dimension. These hidden state vectors are fused with the features in the spatial dimension to generate a spatio-temporal fusion feature matrix. The spatio-temporal fusion feature matrix integrates information in both spatial and temporal dimensions, providing a comprehensive spatio-temporal feature representation for subsequent prediction.

[0138] S34, based on the spatio-temporal fusion feature matrix, predicting the water level, flow and split ratio change quantity of multiple future time periods through a fully connected layer decoder to obtain short-term flood prediction results.

[0139] Specifically, the role of the fully connected layer decoder is to convert the feature matrix that integrates spatio-temporal information into specific prediction results. The decoder is composed of multiple fully connected layers, each containing a certain number of neurons. Each neuron of the fully connected layer is connected to all neurons of the previous layer, and performs feature transformation through weighted summation and a nonlinear activation function (such as ReLU or Sigmoid). In the prediction process, the spatio-temporal fusion feature matrix is flattened or reshaped into a vector form suitable for input to the fully connected layer. The fully connected layer decoder maps the input spatio-temporal feature vector to the output prediction value, i.e. the water level, flow and split ratio change quantity of multiple future time periods, through the learned weight matrix. These prediction values are processed by a suitable activation function to obtain the final short-term flood prediction results, providing key information for flood control decisions.

[0140] where the learning process of the weight matrix is as follows:

[0141] 1) Initialization: Before training begins, the weight matrix is randomly initialized. These weights will be continuously adjusted through the training process to minimize the difference between the prediction value and the true value.

[0142] 2) Forward propagation: During training, the spatio-temporal fused feature matrix is inputted into a fully connected layer decoder. Each neuron in the decoder is connected to each element of the input feature vector, with a linear combination through a weight matrix. For example, assuming the dimension of the input feature vector is D and the dimension of the output prediction is P, the shape of the weight matrix is D x P. Each output prediction is the dot product of the input feature vector and the corresponding column of the weight matrix, plus a bias term.

[0143] 3) Activation function: The result of the linear combination can be introduced nonlinearity through an activation function (such as ReLU or Sigmoid), allowing the model to capture complex nonlinear relationships. For example, the ReLU function sets all negative values to zero while leaving positive values unchanged, helping the model learn more complex feature representations.

[0144] 4) Loss calculation: The difference between the predicted value and the true value is quantified by a loss function (such as mean squared error). The value of the loss function reflects the current prediction performance of the model.

[0145] 5) Backpropagation: Through the backpropagation algorithm, the gradient of the loss function with respect to the weight matrix is calculated. This involves calculating the gradient from the output layer, layer by layer, to the input layer. The gradient represents the degree of contribution of each weight to the loss function.

[0146] 6) Weight update: Based on the calculated gradient, the weight matrix is updated using an optimization algorithm (such as stochastic gradient descent or Adam optimizer). The optimization algorithm adjusts the weights to reduce the value of the loss function. For example, in stochastic gradient descent, the formula for weight update is where η is the learning rate, is the gradient of the loss function with respect to the weight, w new is the updated weight matrix, w old is the weight matrix before updating.

[0147] 7) Iterative training: The above process is repeated for multiple iterations (epochs) on the training dataset until the prediction performance of the model reaches a satisfactory level or the loss function converges. As training progresses, the weight matrix gradually adjusts to a state that best maps the input features to the output predictions.

[0148] S35, for the diversion nodes in the river network, the flow distribution ratio of each branch river is calculated based on the predicted diversion ratio change; according to the flow distribution ratio and the short-term flood prediction result, the flood propagation path and the dynamic distribution data of the diversion ratio are generated combined with the topological relationship of the upstream and downstream nodes.

[0149] Specifically, in a river network, flood is distributed to different branch channels at a certain proportion at a bifurcation node. The predicted bifurcation ratio change in step S34 provides the flow distribution proportion of each bifurcation node in the future period. For example, for a bifurcation node, the predicted bifurcation ratio can be 30% and 70%, indicating that the flood will be distributed to two branch channels in these two proportions respectively.

[0150] When generating the flood propagation path and bifurcation ratio dynamic distribution data, first, the flow distribution proportion of each branch channel is calculated according to the predicted bifurcation ratio. The calculation of the bifurcation ratio is based on the predicted flow change and the topology of the river network. For example, at a bifurcation node, if the predicted total flow is Q, and the bifurcation ratios are r1 and r2, then the flows of the two branch channels are Qxr1 and Qxr2 respectively.

[0151] Next, the propagation path of the flood in the river network is simulated in combination with the short-term flood prediction results and the topological relationship of the upstream and downstream nodes. The predicted water level, flow and bifurcation ratio change can be used to determine the propagation direction and speed of the flood in each node and river section in combination with the connection relationship of the river network.

[0152] Finally, these data are integrated and visualized to generate the flood propagation path and bifurcation ratio dynamic distribution data. These data can be used to guide the construction and operation of flood control projects, such as determining the embankment section that needs to be strengthened, reasonably scheduling the discharge flow of the reservoir, etc. At the same time, it also provides decision support for emergency management such as resident evacuation and material allocation, effectively reducing the loss caused by flood disasters.

[0153] In this embodiment, the spatial convolution module and the time convolution module of the flood dynamic routing prediction model are used to extract and fuse the flood features from the spatial and temporal dimensions respectively. The spatial convolution module accurately propagates hydrological state information through graph convolution operation, and the time convolution module captures the subtle evolution law of hydrological sequence by means of gated recurrent unit. The two cooperate to generate a spatio-temporal fusion feature matrix, greatly improving the model's perception ability of flood dynamics. Finally, the spatio-temporal fusion features are efficiently decoded by the fully connected layer decoder to realize accurate prediction of the water level, flow and bifurcation ratio change in multiple future periods, and the flow distribution proportion of each branch channel is accurately calculated based on the bifurcation ratio change to generate detailed flood propagation path and bifurcation ratio dynamic distribution data, which provides time-efficient and accurate guidance information for flood rescue work, and helps relevant departments to deploy targeted flood control measures in advance, improving the efficiency of flood disaster prevention and control.

[0154] In an alternative embodiment, based on the flood propagation path and bifurcation ratio dynamic distribution data, a flood inundation simulation is performed in combination with a digital elevation model to generate a flood control scheduling scheme and risk warning information, including the following steps:

[0155] S41. Based on the predicted water level changes at each node in the future time period along the flood propagation path, and combined with the raster elevation data of the digital elevation model, calculate the area where the water level exceeds the ground elevation and generate an initial inundation range distribution map.

[0156] Specifically, a DEM provides detailed information on surface elevation within the study area, stored in raster format. Each raster cell corresponds to a small region of the surface and stores the elevation values ​​for that region. Before use, the DEM data undergoes preprocessing, including missing value imputation, outlier handling, and resolution adjustment. For example, if the original DEM has a resolution of 30 meters, but the study requires higher precision, interpolation algorithms (such as Kriging or spline interpolation) can be used to upscale it to 10 meters or higher. Preprocessed DEM data can more accurately reflect subtle changes in surface elevation. Water level changes at various nodes over future periods are obtained from flood dynamics prediction models. This data reflects water level information at key points as floods propagate through the river network.

[0157] The predicted water level change is compared point-by-point with the DEM raster elevation data. For each raster cell, the difference between its elevation and the predicted water level is calculated. If the predicted water level is higher than the raster cell's elevation, the cell is considered a potential flood zone. For example, if the DEM elevation of a certain area is 30 meters and the predicted flood level is 32 meters, then that raster cell will be marked as a flood zone.

[0158] Further calculations are made for the inundation depth, which is the difference between the predicted water level and the elevation of the grid cells. Inundation depth is used to assess flood risk; shallow inundation (e.g., 0-1 meter) may have a greater impact on agricultural land, while deep inundation (e.g., 3 meters or more) may cause severe damage to residential and industrial areas. For example, an inundation depth of 2 meters for a grid cell indicates that the area will experience moderate inundation, potentially causing some damage to buildings and infrastructure.

[0159] Threshold analysis is used to determine the boundaries of areas with different inundation depths. Spatial analysis tools in GIS software, such as "raster calculation" and "regional statistics," can be used to identify continuous inundation areas. For example, areas with an inundation depth greater than 0.5 meters can be defined as effective inundation zones, and the software will automatically identify and extract these areas. The selection of thresholds should be based on actual flood risk assessment standards, which can be provided by local water resources departments or relevant regulations.

[0160] The identified inundation areas are visualized to generate an initial inundation distribution map. Different colors can be used to represent different inundation depths on the map, such as light blue for shallow inundation (0-1 meter), medium blue for moderate inundation (1-3 meters), and dark blue for deep inundation (over 3 meters). Map elements such as legends, scale bars, and coordinate systems are added to ensure the completeness and readability of the distribution map.

[0161] S42, superimpose the river boundary constraint condition in the directed river network topology graph on the initial inundation range distribution graph to generate an inundation area vector boundary; wherein the river boundary constraint condition comprises dike elevation data and water conservancy building position information.

[0162] Specifically, the dike is an important flood control engineering facility, and its elevation data is crucial for determining whether the flood will overtop the dike. The dike elevation data can be obtained by field measurement or from relevant water conservancy departments. In the analysis, if the predicted water level is lower than the dike elevation, the area inside the dike is relatively safe; if the predicted water level exceeds the dike elevation, the area outside the dike may be flooded. For example, the elevation of a certain dike is 35 meters, and the predicted flood water level is 36 meters, then the area outside the dike will be marked as a flooded area.

[0163] Water conservancy buildings such as reservoirs, sluices, and pumping stations have a significant impact on flood regulation and propagation. Their position information can be obtained through the geographic coordinate system (GPS) and stored in GIS software in the form of points, lines, or surfaces. For example, the position information of a certain reservoir includes its dam site coordinates, reservoir capacity curve, and flood discharge capacity, etc. These information is used to evaluate the regulating effect of the reservoir on downstream flood.

[0164] Superimpose the dike elevation data and water conservancy building position information on the initial inundation range distribution graph. In GIS software, the "spatial overlay analysis" tool can be used to overlay the vector format constraint condition layer with the raster format inundation range layer. For example, after superimposing the dike elevation data with the initial inundation range distribution graph, the area within the protection range of the dike can be identified and excluded from the flooded area.

[0165] According to the superposition result, correct the initial inundation range. For example, if an area is within the protection range of the dike and the dike is not overtopped, even if the DEM shows that the elevation of the area is lower than the predicted water level, it will be excluded from the flooded area. The corrected inundation range is more consistent with the actual situation, improving the accuracy of the analysis.

[0166] Convert the corrected inundation range raster data into vector boundaries. In GIS software, the "raster to polygon" tool can be used to convert raster cells into vector polygons. Vector boundaries represent geographical entities in the form of points, lines, and surfaces, with higher precision and spatial analysis capabilities, facilitating integration and analysis with other geographical data.

[0167] Optimize the generated vector boundaries, such as smoothing the boundaries and removing small area patches. The "boundary smoothing" and "area screening" tools can be used to improve the aesthetics and practicality of the vector boundaries. For example, smoothing the boundaries can remove jagged boundaries caused by data precision issues, making the outline of the flooded area smoother and more natural.

[0168] S43, based on the dynamic distribution data of the split ratio, combined with the vector boundary of the flooded area, the economic loss and personnel impact under different flooding depths are evaluated, and the zoning flood risk level assessment map is generated.

[0169] Specifically, based on land use data and asset distribution data, the economic loss in the flooded area is evaluated. Land use data can be obtained from remote sensing images or land use planning maps, and asset distribution data can be obtained through field investigation or statistical yearbook. The specific calculation formula is as follows:

[0170]

[0171] Where, asset value i represents the total value of the i-th asset, and the loss rate i represents the expected loss proportion of the i-th asset in the flood. For example, the asset value of industrial land in a certain area is 10 million yuan, and the loss rate is 0.6, then the economic loss of the area is 600 million yuan.

[0172] Combined with population distribution data and emergency response capability, the personnel impact in the flooded area is evaluated. Population distribution data can be obtained from population census data or mobile phone signaling data, and emergency response capability can be obtained from local emergency management departments. The specific calculation formula is as follows:

[0173]

[0174] Where, population density x represents the population density of the j-th grid unit, affected area j represents the flooded area of the j-th grid unit, and emergency response coefficient j represents the emergency response capability of the j-th grid unit. For example, the population density of a certain area is 100 people per square kilometer, the affected area is 5 square kilometers, and the emergency response coefficient is 0.8, then the personnel impact of the area is 400 people.

[0175] When constructing the risk index system, a risk assessment system containing multiple indexes such as flooding depth, economic loss, and personnel impact is constructed. Different indexes are given corresponding weights, for example, the weight of flooding depth is 0.4, the weight of economic loss is 0.3, and the weight of personnel impact is 0.3. The determination of the weight should be based on expert opinions and statistical analysis of historical flood data to ensure its rationality and scientificity.

[0176] According to the weight and score of each index, the comprehensive risk index is calculated. For example, the flooding depth score of a certain area is 80, the economic loss score is 60, and the personnel impact score is 70, then the comprehensive risk index is: Comprehensive risk index = 0.4 x 80 + 0.3 x 60 + 0.3 x 70 = 71

[0177] According to the comprehensive risk index, the study area is divided into different risk levels, such as low risk (index 0-30), medium risk (31-60), high risk (61-100). The flood risk level assessment map is generated, which provides a direct basis for flood control decision-making. For example, high-risk areas will be given priority protection, medium-risk areas will take appropriate flood control measures, and low-risk areas will take fewer flood control measures.

[0178] S44, according to the flood risk level assessment map, a set of flood control scheduling schemes is generated by a multi-objective optimization algorithm; wherein the multi-objective optimization algorithm minimizes the flood loss, minimizes the scheduling cost and balances the upstream and downstream risks, and the decision variables include reservoir discharge, gate opening and flood diversion time.

[0179] Specifically, a multi-objective optimization model is constructed, which includes minimizing flood loss, minimizing scheduling cost and balancing upstream and downstream risks. For example, the objective function can be represented as:

[0180]

[0181] Wherein, f1 is the flood loss, L i，t is the loss coefficient of region i at time t, D i，t is the flood depth of region i at time t. f2 is the scheduling cost, C j is the cost coefficient of scheduling measure j, U j is the usage degree of scheduling measure j. f3 is the upstream and downstream risk difference, R k is the risk value of upstream region k, R l is the risk value of downstream region l.

[0182] The decision variables include reservoir discharge, gate opening and flood diversion time. For example: the value range of reservoir discharge Q is [Q min , Q max ]. The value range of gate opening H is [0, H max ]. The value range of flood diversion time T is [T start , T end ].

[0183] The optimization process of the multi-objective optimization algorithm is as follows:

[0184] 1) Algorithm selection and parameter setting: select a suitable multi-objective optimization algorithm, such as non-dominated sorting genetic algorithm (NSGA-II). Set the algorithm parameters, such as population size 100, maximum iteration number 500, crossover probability 0.8, mutation probability 0.2. The selection of parameters should be based on the convergence speed and solution quality of the algorithm, and the optimal parameter combination is determined through multiple tests.

[0185] 2) Scheme generation and evaluation: Run multi-objective optimization algorithm to generate multiple flood control scheduling schemes. Each scheme corresponds to a set of decision variable values, and the optimization objective values of each scheme are evaluated through simulation. For example, scheme A corresponds to a reservoir discharge of 100 m / s, a gate opening of 50%, a flood diversion time of 6:00, a inundation loss of 10 million yuan, a scheduling cost of 500 million yuan, and a risk difference between upstream and downstream of 10. The evaluation process should consider the feasibility and effectiveness of different schemes in actual operation to ensure that the generated schemes have practical application value. 3

[0186] 3) Scheme set formation: Select Pareto optimal solutions from multiple schemes to form a set of flood control scheduling schemes. The definition of Pareto optimal solution is: there is no other scheme that is superior to this scheme in all optimization objectives. For example, scheme A and scheme B are both Pareto optimal solutions, scheme A has lower inundation loss but higher scheduling cost, and scheme B has lower scheduling cost but higher inundation loss. The formation of the scheme set should be based on strict screening criteria to ensure that each scheme has unique advantages in different objectives.

[0187] S45, based on the set of flood control scheduling schemes, evaluate the risk probability distribution of different schemes through Monte Carlo simulation; select the Pareto optimal solution as the recommended scheme according to the risk probability distribution, and generate risk warning information corresponding to the recommended scheme; the risk warning information includes warning level, response time and emergency measures.

[0188] Specifically, Monte Carlo simulation is a numerical calculation method through random sampling. For each flood control scheduling scheme, a large number of random simulations are performed to evaluate its performance under different flood scenarios. For example, simulate 1000 different flood events, the flood flow, rainfall and other parameters in each simulation are randomly generated, the random generation of parameters should be based on the statistical distribution of historical flood data, such as using normal distribution or lognormal distribution.

[0189] According to the simulation results, calculate the risk probability distribution of each scheme. For example, scheme A has 100 times of high risk in 1000 simulations, its high risk probability is 10%; scheme B has 50 times of high risk in 1000 simulations, its high risk probability is 5%. The calculation of risk probability should use statistical analysis methods such as frequency analysis or probability density function fitting to ensure the reliability and accuracy of the results.

[0190] ​Compare the risk probability distribution of different schemes, and select the Pareto optimal solution as the recommended scheme. The definition of Pareto optimal solution is: achieve the best balance between multiple optimization objectives, and cannot further improve one objective without deteriorating another objective. For example, scheme C and scheme D are both Pareto optimal solutions, scheme C has lower inundation loss but higher scheduling cost, and scheme D has lower scheduling cost but higher inundation loss. According to actual needs and preferences, choose scheme C as the recommended scheme. In the selection process, the importance of different objectives should be considered comprehensively, such as under the premise of protecting life safety, the scheme with lower inundation loss can be selected first.

[0191] After determining the recommended scheme, conduct comprehensive evaluation and verification to ensure its feasibility and effectiveness in actual application. For example, through actual drills or historical data backtesting, verify the performance of the recommended scheme under different flood scenarios, and timely discover and solve potential problems.

[0192] Generate risk warning information according to the recommended scheme, including warning level, response time and emergency measures. For example: according to risk probability and potential loss, divide into low risk (green), medium risk (yellow), high risk (orange) and very high risk (red). The division of warning level should be based on strict risk assessment standards, such as high risk corresponding to risk probability greater than 50% and potential loss exceeding 10 million yuan. Specify the time requirements for measures under different risk levels, such as high risk needs to be responded within 1 hour, medium risk needs to be responded within 2 hours. The determination of response time should be based on emergency response capacity and flood propagation speed, to ensure that necessary protection measures are completed before the flood arrives. Emergency measures provide specific emergency measure suggestions, such as personnel evacuation, material transfer, dike reinforcement, etc. Emergency measures should be operable and targeted, such as evacuation of residents in high-risk areas, should clearly specify evacuation routes, temporary resettlement sites and responsible persons.

[0193] Release risk warning information through various channels, such as mobile phone text message, radio, television, etc., to ensure that relevant information is timely and accurately conveyed to relevant departments and the public.

[0194] In this embodiment, the flood propagation path and split ratio data are combined with the digital elevation model for depth inundation simulation. Through multi-step refinement processing, the initial inundation range distribution map, the inundation area vector boundary are generated in turn, and the economic loss and personnel impact under different inundation depths are further evaluated, and finally the intuitive zoning flood risk level evaluation map is formed. This process fully considers the influence of the space-time dynamic characteristics of the flood and the topography and other factors, making the inundation simulation result more accurate and detailed. In terms of generating flood control scheduling scheme, this method uses multi-objective optimization algorithm to intelligently generate and screen the scheme, comprehensively considers the inundation loss, scheduling cost and upstream and downstream risk and other factors, and forms a comprehensive and optimized scheme set. At the same time, the Monte Carlo simulation is introduced to rigorously evaluate the risk probability distribution of the scheme, and the Pareto optimal solution is selected as the recommended scheme to ensure the scientificity and rationality of the flood control scheme. Finally, the generated risk warning information contains rich practical content, providing all-round decision support for flood control and disaster reduction work, effectively improving the scientificity and accuracy of flood control decision-making, minimizing the risk of flood disasters, and protecting people's life and property safety and social stability.

[0195] The above-mentioned disaster multi-source situation intelligent perception management and control method collects hydrological and meteorological and terrain data by deploying multi-source monitoring equipment, generates a standardized data set through space-time alignment and anomaly cleaning; based on this, the river network topological relationship is extracted to construct a directed topological graph containing node and edge attributes; then a graph neural network model is designed, and a flood dynamic routing prediction model is trained through a physical constraint loss function embedded with the water balance principle; real-time hydrological data is input into the model for graph convolution operation to predict the water level, flow and split ratio changes of each node in the future period; finally, flood control scheduling scheme and risk warning information are generated through inundation simulation analysis combined with a digital elevation model. This technical solution realizes the deep integration of physical mechanisms and data-driven, effectively captures the flood propagation law under the topological constraints of the river network using graph neural networks, significantly improves the calculation efficiency while ensuring the prediction accuracy, and provides real-time and reliable decision support for flood disaster prevention and control in complex river network areas.

[0196] It should be understood that, although each step in the flowchart involved in each embodiment as described above is shown in sequence according to the direction of the arrow, these steps are not necessarily executed in the order indicated by the arrow. Unless otherwise specified herein, there is no strict order limitation for the execution of these steps, and these steps can be executed in other orders. Moreover, at least part of the steps in the flowchart involved in each embodiment as described above can include multiple steps or stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but can be executed in rotation or alternation with at least part of other steps or steps or stages in other steps.

[0197] Based on the same inventive concept, the embodiments of the present application also provide a system for implementing the disaster multi-source situation intelligent perception management and control method described above. The system provides a solution to the implementation scheme as described in the above method, and therefore the specific limitations in one or more disaster multi-source situation intelligent perception management and control system embodiments provided below can refer to the limitations of the disaster multi-source situation intelligent perception management and control method described above, which will not be repeated here.

[0198] In one exemplary embodiment, as shown in Figure 3 A disaster multi-source situation intelligent perception management and control system 30 is provided for implementing the method in each of the method embodiments described above, which includes:

[0199] A multi-source data acquisition and standardization module 31 is configured to acquire a historical multi-source heterogeneous data set composed of water level, flow, rainfall, and terrain data based on hydrological monitoring sites, weather radar stations, and remote sensing equipment deployed in the river network area, perform spatio-temporal alignment and outlier cleaning processing on the historical multi-source heterogeneous data set, and generate a standardized spatio-temporal data set.

[0200] A river network topology modeling module 32 is configured to extract river channel topology relationships through hydrological analysis based on the standardized spatio-temporal data set, and construct a directed river network topology graph including node attribute vectors and edge attribute vectors according to the river channel topology relationships; wherein the nodes in the directed river network topology graph correspond to river junctions, and the edges in the directed river network topology graph correspond to river sections.

[0201] A graph neural network training module 33 is configured to construct a graph neural network model based on the directed river network topology graph, train the graph neural network model through a physical constraint loss function, and obtain a flood dynamic routing prediction model.

[0202] A flood dynamic prediction module 34 is configured to input real-time hydro-meteorological data of each node in the directed river network topology graph into the flood dynamic routing prediction model, and predict water levels, flows, and diversion ratios of each node in a future time period through graph convolution operation to obtain flood propagation path and diversion ratio dynamic distribution data.

[0203] A flooding simulation and dispatching decision module 35 is configured to perform flooding simulation based on the flood propagation path and diversion ratio dynamic distribution data in combination with a digital elevation model to generate a flood control dispatching scheme and risk warning information.

[0204] The embodiments of the present application also provide a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps in each of the method embodiments described above when executing the computer program.

[0205] The embodiments of the present application further provide a computer readable storage medium, which stores a computer program. The computer program is executed by a processor to implement the steps in the above-mentioned method embodiments.

[0206] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts are described in the part of the method embodiments. The above-described device embodiments are merely illustrative, wherein the components described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., can be located in one place or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the present disclosure according to actual needs. Those skilled in the art can understand and implement it without creative labor.

[0207] The above-described embodiments only express several implementation manners of the embodiments of the present application, and the description is relatively specific and detailed, but it should not be understood as a limitation on the patent scope of the embodiments of the present application. It should be pointed out that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the embodiments of the present application, and these all belong to the protection scope of the embodiments of the present application.

Claims

1. A method for intelligent perception and control of multi-source disaster situations, characterized in that, The method includes: S1. Based on hydrological monitoring stations, meteorological radar stations and remote sensing equipment deployed in the river network area, collect historical multi-source heterogeneous datasets consisting of water level, flow rate, rainfall and topographic data, perform spatiotemporal alignment and outlier cleaning on the historical multi-source heterogeneous datasets, and generate standardized spatiotemporal datasets. S2. Based on the standardized spatiotemporal dataset, the river channel topology is extracted through hydrological analysis, and a directed river network topology graph containing node attribute vectors and edge attribute vectors is constructed according to the river channel topology. The nodes in the directed river network topology graph correspond to river confluence points, and the edges in the directed river network topology graph correspond to river segments. S3. Based on the directed river network topology, construct a graph neural network model; train the graph neural network model using a physical constraint loss function to obtain a flood dynamic routing prediction model; S4. Input the real-time hydrological and meteorological data of each node in the directed river network topology graph into the flood dynamic routing prediction model, and predict the water level, flow rate and diversion ratio of each node in the future period through graph convolution operation to obtain the dynamic distribution data of flood propagation path and diversion ratio. S5. Based on the dynamic distribution data of the flood propagation path and the diversion ratio, and combined with the digital elevation model, flood simulation is performed to generate flood control scheduling plan and risk warning information.

2. The method according to claim 1, characterized in that, The construction of a graph neural network model based on the directed river network topology includes: S11. Using the original node feature matrix of the directed river network topology as input, and combining it with the adjacency matrix of the directed river network topology, construct a spatial convolution module. The spatial convolution module is used to process the topological relationships in the river network topology graph, and the output of the spatial convolution module is a spatial feature matrix for each time dimension; if there is a downstream river segment connection from node a to node b, then element A in the adjacency matrix... ab =1, otherwise A ab =0; the original node feature matrix represents the hydrological characteristics of each node; The spatial convolution module performs graph convolution operations based on Chebyshev polynomial approximation, and the expression for the graph convolution operation is: in, Let X represent the spatial characteristic matrix at time t. t T represents the original node feature matrix at time t; k (·) is a Chebyshev polynomial of order k, θ k T represents the trainable parameters of order k, where K is the order of the polynomial; k (·) is defined as: T0(x)=1, T1(x)=x, T k (x)=2xT k-1 (x)-T k-2 (x); Where L is the Laplace matrix, representing the spatial correlation strength between river network nodes in the directed river network topology graph, and the formula for calculating the Laplace matrix is ​​L = ID. -1 / 2 AD -1 / 2 A is the adjacency matrix, and D is the degree matrix; the diagonal elements of the degree matrix D ii =∑ j A ij λ represents the out-degree of node i in the directed river network topology graph, where I is the identity matrix; max The largest eigenvalue of the Laplacian matrix; S12. Using the spatial feature matrix as input, construct a temporal convolution module; the output of the temporal convolution module is the hidden state time series; The temporal convolution module employs a gated recurrent unit structure, and its operation process is as follows: Where, r t To reset the door, z t To update the door, As a candidate state, h t For the final state, h t-1 W is the hidden state of the previous time step. r W z W h Let σ be the trainable parameter matrix, σ be the sigmoid activation function, and ⊙ be the Hadamard product. S13. Construct a graph neural network framework based on the spatial convolution module and the temporal convolution module to obtain an initial neural network; wherein, the output of the initial neural network is a three-dimensional spatiotemporal fusion feature tensor obtained by stacking the hidden state time series along the time dimension.

3. The method according to claim 2, characterized in that, The process of training the graph neural network model using a physical constraint loss function to obtain a flood dynamic routing prediction model includes: S21. Based on the principle of water balance, construct the continuity equation constraint terms. The expression for the continuity equation constraint terms is: Where M1 represents the set of upstream nodes flowing into node n in the directed river network topology graph, M2 represents the set of downstream nodes of node n in the directed river network topology graph, and Q t (i,n) represents the predicted flow rate from node i to node n at time t, Q t (n,j) represents the predicted flow from node n to node j at time t, A t (n) represents the predicted cross-sectional area of ​​the water flow at node n at time t, Δt represents the time step, and Δx n q represents the average length of all river segments connected to node n. t (n) represents the predicted lateral inflow of the river segment where node n is located, N represents the total number of nodes in the directed river network topology graph, and T represents the number of steps in the total prediction period. S22. The prediction error term of the initial neural network is combined with the constraint term of the continuity equation to form a comprehensive loss function; the expression of the comprehensive loss function is: Where α and β are adjustment coefficients, y t,n Let be the measured flow rate of node n at time t. Let N be the predicted flow value of node n at time t for the initial neural network, and N be the total number of nodes; S23. Extract historical flood event dataset from the standardized spatiotemporal dataset; based on the historical flood event dataset, train the initial neural network using the comprehensive loss function to obtain the flood dynamic routing prediction model.

4. The method according to claim 2, characterized in that, The process involves inputting real-time hydrological and meteorological data of each node in the directed river network topology graph into the flood dynamic routing prediction model. Through graph convolution operations, the model predicts the water level, flow rate, and diversion ratio of each node in future time periods, obtaining dynamic distribution data of flood propagation paths and diversion ratios. This includes: S31. Based on the real-time water level monitoring data and real-time flow monitoring data of each node in the directed river network topology diagram, and combined with the real-time rainfall distribution data obtained by meteorological radar, the data is processed for format standardization to generate a real-time input feature matrix. S32. Input the real-time input feature matrix into the spatial convolution module of the flood dynamic routing prediction model, and propagate hydrological state information on the river network topology through graph convolution operations to generate a node spatial feature matrix. S33. Input the node spatial feature matrix into the temporal convolution module of the flood dynamic routing prediction model, and capture the evolution law of hydrological time series through the gated recurrent unit to generate a spatiotemporal fusion feature matrix. S34. Based on the spatiotemporal fusion feature matrix, the changes in water level, flow rate and diversion ratio in multiple future time periods are predicted by the fully connected layer decoder to obtain short-term flood prediction results. S35. For the diversion nodes in the river network, calculate the flow distribution ratio of each branch channel based on the predicted change in the diversion ratio; and generate dynamic distribution data of flood propagation path and diversion ratio based on the flow distribution ratio and the short-term flood prediction results, combined with the topological relationship of upstream and downstream nodes.

5. The method according to any one of claims 1 to 4, characterized in that, The process involves using the dynamic distribution data of the flood propagation path and the diversion ratio, combined with a digital elevation model, to perform inundation simulation and generate flood control scheduling plans and risk warning information, including: S41. Based on the predicted water level changes at each node in the future time period in the flood propagation path, and combined with the raster elevation data of the digital elevation model, calculate the area where the water level exceeds the ground elevation and generate an initial flood range distribution map. S42. Overlay the river boundary constraints in the directed river network topology map onto the initial inundation range distribution map to generate the inundation area vector boundary; wherein, the river boundary constraints include levee elevation data and hydraulic structure location information; S43. Based on the dynamic distribution data of the diversion ratio and combined with the vector boundary of the inundated area, assess the economic losses and human impacts at different inundation depths and generate a zonal flood risk level assessment map. S44. Based on the zonal flood risk level assessment map, a flood control scheduling scheme set is generated through a multi-objective optimization algorithm; wherein, the multi-objective optimization algorithm takes minimizing inundation loss, minimizing scheduling cost and balancing upstream and downstream risks as optimization objectives, and the decision variables include reservoir discharge volume, gate opening degree and flood diversion timing. S45. Based on the flood control scheduling scheme set, evaluate the risk probability distribution of different schemes through Monte Carlo simulation; select the Pareto optimal solution as the recommended scheme according to the risk probability distribution, and generate the risk warning information corresponding to the recommended scheme; the risk warning information includes the warning level, response time and emergency measures.

6. A multi-source disaster situational intelligent perception and control system, used to implement the method according to any one of claims 1 to 5, characterized in that, The system includes: The multi-source data acquisition and standardization module is used to collect historical multi-source heterogeneous datasets consisting of water level, flow, rainfall, and topographic data based on hydrological monitoring stations, meteorological radar stations, and remote sensing equipment deployed in the river network area. The module performs spatiotemporal alignment and outlier cleaning on the historical multi-source heterogeneous datasets to generate standardized spatiotemporal datasets. The river network topology modeling module is used to extract river channel topological relationships through hydrological analysis based on the standardized spatiotemporal dataset, and construct a directed river network topology graph containing node attribute vectors and edge attribute vectors according to the river channel topological relationships; wherein, the nodes in the directed river network topology graph correspond to river confluence points, and the edges in the directed river network topology graph correspond to river segments; The graph neural network training module is used to construct a graph neural network model based on the directed river network topology graph; and to train the graph neural network model through a physical constraint loss function to obtain a flood dynamic routing prediction model. The flood dynamic prediction module is used to input the real-time hydrological and meteorological data of each node in the directed river network topology map into the flood dynamic routing prediction model, and predict the water level, flow rate and diversion ratio of each node in the future period through graph convolution operation to obtain the dynamic distribution data of flood propagation path and diversion ratio. The flood simulation and scheduling decision module is used to simulate flooding based on the dynamic distribution data of the flood propagation path and the diversion ratio, combined with the digital elevation model, to generate flood control scheduling plans and risk warning information.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 5.

Images