Flood forecasting method, system, storage medium and computing device

Abstract

The present invention relates to the field of flood forecasting and discloses a flood forecasting method, system, storage medium, and computing device. The method comprises: topological modeling, using the river catchment area surrounded by the watershed and the hydrological stations at the catchment area outlet as the nodes of the graph, and constructing edge connections based on the actual flow direction of the river and the connectivity between its tributaries, thereby constructing topological graph structure data that is compatible with hydrological mechanisms; converting month number features to generate dynamic periodic timestamp vectors, splicing them with the topological graph structure data, and inputting them into attention neural network training; the attention neural network architecture includes two independent feature channels, one for processing historical data and the other for processing forecast data. The network supports a variable modular architecture, and after the model is trained, it can output flood forecast results as needed. The present invention improves the accuracy and interpretability of flood prediction, realizing a high-precision, interpretable spatiotemporal joint flood prediction solution.

Description

Technical Field

[0001] The present invention relates to the field of flood forecasting technology, and in particular to a flood forecasting method, system, storage medium and computing device of a topological modeling-time-space coupled attention neural network. Background Art

[0002] Floods are one of the most destructive natural disasters in the world. They can cause economic losses, casualties and ecological damage. Timely and accurate flood forecasts can provide flood prevention decisions and reduce losses.

[0003] The mainstream flood prediction methods include hydrological mechanism models and data-driven models. The mechanism models infer the evolution of rivers within the basin by simulating hydrological processes, while the data-driven models directly mine spatiotemporal laws from the data without the need for explicit physical equations. With the development of sensor technology, satellite remote sensing and hydrological stations, relevant meteorological and hydrological data are becoming increasingly abundant, and data-driven models are developing rapidly.

[0004] Most existing data-driven methods rely solely on time series data for prediction, ignoring the spatial structure of the watershed. CN111915058A provides a flood detection method implemented using a long short-term memory network and transfer learning, as well as a BiGRU multi-step prediction method, system, and storage medium for flood prediction. These methods primarily capture correlations at the time series level, lacking the ability to capture spatial correlations. Existing methods for capturing spatial correlations using graph neural network models generally use rectangular segmentation or directly use hydrological stations as nodes, without considering natural watersheds or the relationship between main and tributary streams. This results in poor model interpretability. For example, the prediction method proposed in the open document "FloodGNN-GRU: a spatio-temporal graph neural network for flood prediction" uses rectangular segmentation of the watershed to construct a graph structure.

[0005] In summary, existing flood forecasting technologies have obvious deficiencies in capturing spatial correlations in data-driven models. Summary of the Invention

[0006] In response to the above problems, the purpose of the present invention is to provide a flood forecasting method, system, storage medium and computing device based on a topological modeling-spatiotemporal coupled attention neural network, which combines the construction method of the topological graph structure constructed by the hydrological mechanism and the spatiotemporal attention convolutional network model, improves the flood prediction accuracy and interpretability, and realizes a high-precision and interpretable spatiotemporal joint flood prediction solution.

[0007] To achieve the above-mentioned purpose, in the first aspect, the technical solution adopted by the present invention is: a flood forecasting method, which includes: using the river catchment area surrounded by the watershed line and the hydrological stations at the catchment outlet as nodes of the topological map, and constructing the connection relationship of the topological map based on the actual flow direction of the river and the connection relationship of each tributary, so that the organizational form of the topological map structure is adapted to the hydrological mechanism; converting the monthly features, generating a dynamic periodic timestamp vector based on the bidirectional embedding of the sine function and the cosine function, splicing it with the topological map structure data to generate spatiotemporal coupling features; inputting the spatiotemporal coupling feature data into the spatiotemporal graph attention convolutional network training with dual independent feature channels; the two independent feature channels in the spatiotemporal graph attention neural network architecture process historical data and forecast data respectively, and the network supports a variable modular architecture; the meteorological data and historical flow of a certain basin during the forecast period are topologically constructed and spatiotemporally coupled with the meteorological data to form data of the same type as the training model, and inputting the trained spatiotemporal attention convolutional network to obtain the flood forecast result for the basin.

[0008] Furthermore, the spatiotemporal graph attention convolutional network with dual independent feature channels includes multiple STACNN modules connected in series;

[0009] Each STACNN module includes: temporal attention module, spatial attention module, spatial convolution module, temporal convolution module, residual connection module and output connection module;

[0010] The input data is transmitted to the temporal attention module and the residual connection module respectively. After the temporal attention module assigns different weights to different graph structure data in the time dimension, it is multiplied with the input data and transmitted to the spatial attention module. Among them, the input data of the first-level STACNN module is the encoded graph structure data, and the input data of each subsequent STACNN module is the output data of the previous level STACNN module.

[0011] The spatial attention module assigns different weights to the received data in the spatial dimension, multiplies it with the input data, and then processes it through the spatial convolution module and the temporal convolution module in sequence. It is then added to the data output by the residual connection module and output through the output connection module as the input data of the next level STACNN module.

[0012] Furthermore, the temporal attention module is:

[0013]

[0014] in, It is the input data of the STACNN module; To pass Sigmoid The time attention score matrix after function normalization; is the output data after being processed by the temporal attention module.

[0015] Furthermore, the spatial attention module is:

[0016]

[0017] in, is the output data of the spatial attention module; is through Softmax The spatial attention score matrix after function normalization; S Is the module output data; is a learnable parameter; Used to extract features in the time dimension, Used to transform the matrix dimension, mapping it from the channel dimension to the time dimension, Used to extract features in the channel dimension.

[0018] Furthermore, the spatial convolution module adopts a graph convolution operation based on the simplification of Chebyshev polynomials to combine the topological structure of the graph with the node features, obtains neighbor node information through message passing, aggregates the obtained information, updates the aggregation results, and finally obtains the complex structured information in the graph.

[0019] Furthermore, the spatiotemporal attention convolutional network adopts the mean square error (MSE) as the loss function.

[0020] Furthermore, it also includes a method for evaluating the predictive ability of spatiotemporal attention convolutional networks:

[0021] The Nash efficiency coefficient NSE is used to evaluate the prediction results of the hydrological process;

[0022] NSE = 1, indicating that the model simulation values ​​are completely consistent with the observed values ​​and the model prediction effect is perfect;

[0023] NSE = 0, indicating that the model prediction effect is the same as directly using the average of the observed values ​​as the predicted value;

[0024] NSE < 0 means that the model prediction effect is worse than directly using the mean of the observed values ​​as the predicted value.

[0025] In the second aspect, the technical solution adopted by the present invention is: a flood forecasting system, which includes: a topological modeling module, which uses the river catchment area surrounded by the watershed line and the hydrological stations at the catchment outlet as the nodes of the topological map, and constructs the connection relationship of the topological map based on the actual flow direction of the river and the connectivity of each tributary, so that the organizational form of the topological map structure is adapted to the hydrological mechanism; a spatiotemporal coupling module, which converts monthly features, generates a dynamic periodic timestamp vector based on the bidirectional embedding of sine function and cosine function, and splices it with the topological map structure data to generate spatiotemporal coupling features; a spatiotemporal attention convolutional network module, which inputs the spatiotemporal coupling feature data into the spatiotemporal graph attention convolutional network training with dual independent feature channels; the two independent feature channels in the spatiotemporal graph attention neural network architecture process historical data and forecast data respectively, and the network supports a variable modular architecture; a prediction module, which uses topological construction and spatiotemporal coupling to form data of the same type as the training model for the meteorological data and historical flow during the forecast period of a certain river basin, and inputs the trained spatiotemporal attention convolutional network to obtain the flood forecast results for the river basin.

[0026] In a third aspect, the technical solution adopted by the present invention is: a computer-readable storage medium storing one or more programs, wherein the one or more programs include instructions, and when the instructions are executed by a computing device, the computing device executes any one of the above methods.

[0027] In a fourth aspect, the technical solution adopted by the present invention is: a computing device, comprising: one or more processors, a memory and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include instructions for executing any of the above methods.

[0028] Due to the adoption of the above technical solution, the present invention has the following advantages: the present invention combines the construction method of the topological graph structure constructed by the hydrological mechanism and the spatiotemporal attention convolutional network model, which improves the accuracy and interpretability of flood prediction and realizes a high-precision and interpretable spatiotemporal joint flood prediction solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is an overall flow chart of a flood forecasting method according to an embodiment of the present invention;

[0030] Figure 2 2 is a schematic diagram of the structure of the spatiotemporal attention convolutional network in an embodiment of the present invention. DETAILED DESCRIPTION

[0031] In order to overcome the shortcomings of existing data-driven time-series flood prediction models and improve the accuracy and interpretability of flood prediction, the present invention provides a flood forecasting method, system, medium and equipment based on graph structure and spatiotemporal attention convolutional network, and sets up a spatiotemporal attention mechanism to jointly model complex spatiotemporal dependencies.

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field are within the scope of protection of the present invention.

[0033] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0034] In one embodiment of the present invention, Figure 1 As shown, a flood forecasting method is provided, which processes various features in the basin into a graph structure and uses a data-driven spatiotemporal attention convolutional network to forecast river water level and flow. In this embodiment, the method includes the following steps:

[0035] The first step is to construct a graph structure using topological modeling. The nodes of the graph are the river catchments enclosed by the watershed lines and the hydrological stations at their outlets. Each node is associated with the flow and water level characteristics of the hydrological station. The weighted average and cumulative values ​​of meteorological characteristics are calculated by linking the corresponding catchment area with the overlapping area of ​​the reanalysis grid generated based on satellite remote sensing data. The edges of the graph are connected based on the actual flow direction of the river and the connectivity between its tributaries, so that the organization of the graph structure is consistent with hydrological mechanisms. For catchments without outlet hydrological stations, their downstream nodes along the river are merged to ensure that the characteristics of all nodes are preserved.

[0036] The second step is spatiotemporal coupling. The features of the month in which the data is located are extracted. Based on the bidirectional embedding of the sine function and cosine function, a dynamic periodic timestamp vector is generated, which is spliced ​​with the topological structure data to generate spatiotemporal coupling features.

[0037] In the third step, the spatiotemporal coupled feature data is input into the spatiotemporal graph attention convolutional network (DFC-STAGCN) with dual independent feature channels for training; the two independent feature channels in the spatiotemporal graph attention neural network architecture process historical data and forecast data respectively, and the network supports a variable modular architecture.

[0038] In this embodiment, Figure 2 As shown in the figure, a dual-independent feature channel spatiotemporal graph attention convolutional network (DFC-STAGCN) is constructed, which contains multiple cascaded spatiotemporal attention convolutional networks, namely STACNN modules. Each STACNN module is composed of a temporal attention module, a spatial attention module, a spatial convolution module, a temporal convolution module, a residual connection module, and an output connection module:

[0039] (1) Temporal attention module: In the temporal dimension, there is a correlation between different features in the previous and subsequent time periods. For example, rainfall in the previous period will affect the flow rate of the measuring station and the soil water content of the corresponding catchment area. The correlation is different in different situations. The use of the attention mechanism can adjust the different weights between different data through training. The corresponding formula is as follows:

[0040]

[0041] in, It is the input data of the STACNN module; is a learnable parameter; Used to extract features on the node dimension, Used to transform the matrix dimension, mapping it from channel dimension to node dimension, Used to extract features in the channel dimension; Sigmoid The activation function is used to limit the range of results; It is the time attention score matrix normalized by the Sigmoid function.

[0042] (2) Spatial attention module: In the graph constructed based on catchment areas and hydrological stations, all nodes are part of the same interconnected watershed, which leads to the different characteristics of different nodes affecting each other. Considering that different nodes have different geographical locations, shapes, meteorological conditions, etc., we have a high degree of confidence that this influence is not uniform. The attention mechanism can be used to automatically capture the influence of different weights between different nodes through training. The corresponding formula is as follows:

[0043]

[0044] in, It is the input data of the STACNN module; is the input data after the temporal attention module; is a learnable parameter; Used to extract features in the time dimension, Used to transform the matrix dimension, mapping it from the channel dimension to the time dimension, Used to extract features in the channel dimension; Sigmoid The activation function is used to limit the range of results; is through Softmax The spatial attention score matrix after function normalization.

[0045] Multiply the spatial attention matrix output by the module with the original data to obtain the data after the spatial attention module .

[0046] (3) Spatial convolution module: The graph convolution network migrates the convolution operation from the image domain to the graph domain. It can operate directly on irregular graph structures without converting them into regular grid data, thereby obtaining complex structured information in the graph. It combines the topological structure of the graph with node features, obtains neighbor node information through message passing, aggregates information, updates aggregation results, and learns to expand the node field of view through multi-layer stacking. In order to reduce the computational difficulty and simplify the calculation, this module uses a graph convolution operation based on the simplification of Chebyshev polynomials:

[0047]

[0048] in, is the graph convolution kernel, are the polynomial coefficients, Represents graph convolution operation; is a Chebyshev polynomial, recursively defined as ,in . K is the order of the Chebyshev polynomial; is the normalized Laplacian matrix.

[0049] in:

[0050] .

[0051] (4) Temporal convolution module: The temporal convolution module is located at the end of the entire module to ensure that all nodes can obtain the corresponding spatial dependencies:

[0052]

[0053] in, are the parameters of the temporal convolution kernel; Represents a standard convolution operation; ReLU is the activation function.

[0054] (5) Residual connection module: Residual connection avoids the gradient vanishing or exploding problem by adding the input directly to the output and accelerates the training of the model. If the dimensions of the input and output are different, an additional convolutional layer is required to adjust them:

[0055]

[0056] in, is the convolution kernel parameter; is the bias parameter.

[0057] (6) Output connection module: The output connection module includes a RELU activation function and a layer normalization module, which can provide nonlinear transformation and stability for the previous module, and alleviate the gradient disappearance or gradient explosion problem.

[0058] In the fourth step, the meteorological data and historical flow of a certain river basin during the forecast period are combined with the meteorological data through the topological construction and spatiotemporal coupling of the first and second steps to construct spatiotemporal graph structure data of the same type as the training model. This data is input into the spatiotemporal attention convolutional network constructed in the third step as training data, and the neural network training is performed to obtain the flood forecast model for the river basin.

[0059] In the above embodiments, the spatiotemporal attention convolutional network uses the mean square error (MSE) as the loss function.

[0060] In each of the above embodiments, after the fourth step is completed, a step of evaluating the prediction ability of the spatiotemporal attention convolutional network is also included. Specifically, the evaluation method is to use the Nash efficiency coefficient NSE to evaluate the prediction results of the hydrological process. Specifically:

[0061] NSE = 1, indicating that the model simulation values ​​are completely consistent with the observed values ​​and the model prediction effect is perfect;

[0062] NSE = 0, indicating that the model prediction effect is the same as directly using the average of the observed values ​​as the predicted value;

[0063] NSE < 0 means that the model prediction effect is worse than directly using the mean of the observed values ​​as the predicted value.

[0064] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0065] (1) Establish a training dataset. First, the reanalyzed grid meteorological data and the watersheds divided by the watershed are unified into the same geographic projection coordinate system WGS84. The grid meteorological data are weighted averaged or weighted summed according to the corresponding area of ​​the watershed to obtain the meteorological data characteristics of each watershed corresponding to its own location and area. Each watershed is associated with the flow and water level characteristics of its outlet hydrological station. For watersheds without outlet hydrological stations, they are merged along the river distribution to the downstream nodes to ensure that the characteristics of all nodes remain complete. In this way, the nodes of the graph structure are constructed. The connection relationship is constructed by the actual flow direction of the rivers between different watersheds and the connection relationship between each tributary. The edge connection relationship of the graph is constructed. For the constructed dataset, missing value filling and data standardization are performed. The prediction time window size of the training data is set according to the required prediction time length, device storage capacity and device computing power. The training data is processed into the data format required by the spatiotemporal attention convolution model.

[0066] (2) Establish a spatiotemporal graph attention convolutional network with dual independent feature channels based on the STACNN module. Each STACNN module is composed of a temporal attention module, a spatial attention module, a spatial convolution module, a temporal convolution module, a residual connection module, and an output connection module. The overall spatiotemporal attention convolutional neural network consists of a periodic encoding module, several STACNN modules, and a fully connected output module. The specific network structure can be referred to in the accompanying description. Figure 1 Structural diagram.

[0067] (3) Select mean square error MSE as the loss function of the model:

[0068]

[0069] Where n is the number of samples, is the i-th true value, is the i-th predicted value.

[0070] (4) A training method for the training data in (1) and the spatiotemporal attention convolutional neural network (2) is established, and the model parameters are iteratively learned through a graphics card. Each iteration adjusts the model parameters so that the mean square error of the loss function in (3) between the predicted values ​​and the true values ​​generated during the training process is reduced. Through multiple training and iterations, the model parameters that are most suitable for the corresponding watershed are generated.

[0071] (5) Model prediction ability evaluation: The prediction results of the hydrological process are evaluated by the Nash efficiency coefficient NSE. NSE is a statistical indicator used to evaluate the performance of hydrological models or other prediction models. It measures the degree of fit between the model simulation value and the observed value, and its value range is .

[0072] NSE = 1: Indicates that the model simulation value is completely consistent with the observed value, and the model prediction effect is perfect.

[0073] NSE = 0: Indicates that the model prediction effect is equivalent to directly using the mean of the observed values ​​as the prediction value.

[0074] NSE < 0: Indicates that the model's prediction performance is worse than directly using the mean of the observed values ​​as the predicted value.

[0075] The calculation formula of Nash equilibrium efficiency coefficient (NSE) is as follows:

[0076]

[0077] in, represents the observation value at the "i" time step; represents the simulation value of the i-th time step; represents the average value of the observations; n represents the total number of time steps.

[0078] (6) After training and evaluation, the parameters of the spatiotemporal attention convolutional neural network are fixed and saved to generate an inference model for predicting future floods. At the same time, the feature data required for predicting future floods are constructed according to the construction method of the training data. Combining the feature data and the inference model can predict future floods in the basin.

[0079] In one embodiment of the present invention, a flood forecasting system is provided, comprising:

[0080] The topological modeling module uses the river catchment area surrounded by the watershed line and the hydrological stations at the catchment outlet as the nodes of the topological map. The connection relationship of the topological map is constructed based on the actual flow direction of the river and the connection relationship between the tributaries, so that the organizational form of the topological map structure is adapted to the hydrological mechanism.

[0081] The spatiotemporal coupling module converts the monthly features and generates a dynamic periodic timestamp vector based on the bidirectional embedding of the sine and cosine functions. It then combines the vector with the topological structure data to generate spatiotemporal coupling features.

[0082] The spatiotemporal attention convolutional network module inputs spatiotemporal coupled feature data into the spatiotemporal graph attention convolutional network training with two independent feature channels. The two independent feature channels in the spatiotemporal graph attention neural network architecture process historical data and forecast data respectively. The network supports a variable modular architecture.

[0083] The prediction module combines the meteorological data and historical flow of a certain river basin during the forecast period with the meteorological data through topological construction and spatiotemporal coupling to form data of the same type as the training model, and inputs the trained spatiotemporal attention convolutional network to obtain the flood forecast results for the river basin.

[0084] In the above embodiment, the spatiotemporal graph attention convolutional network with dual independent feature channels includes multiple STACNN modules connected in series;

[0085] Each STACNN module includes: temporal attention module, spatial attention module, spatial convolution module, temporal convolution module, residual connection module and output connection module;

[0086] The input data is transmitted to the temporal attention module and the residual connection module respectively. After the temporal attention module assigns different weights to different graph structure data in the time dimension, it is multiplied with the input data and transmitted to the spatial attention module. Among them, the input data of the first-level STACNN module is the encoded graph structure data, and the input data of each subsequent STACNN module is the output data of the previous level STACNN module.

[0087] The spatial attention module assigns different weights to the received data in the spatial dimension, multiplies it with the input data, and then processes it through the spatial convolution module and the temporal convolution module in sequence. It is then added to the data output by the residual connection module and output through the output connection module as the input data of the next level STACNN module.

[0088] In this embodiment, the temporal attention module is:

[0089]

[0090] in, It is the input data of the STACNN module; To pass Sigmoid The time attention score matrix after function normalization; is the output data after being processed by the temporal attention module.

[0091] In this embodiment, the spatial attention module is:

[0092]

[0093] in, is the output data of the spatial attention module; is through Softmax The spatial attention score matrix after function normalization; S Is the module output data; is a learnable parameter; Used to extract features in the time dimension, Used to transform the matrix dimension, mapping it from the channel dimension to the time dimension, Used to extract features in the channel dimension.

[0094] In this embodiment, the spatial convolution module adopts a graph convolution operation based on the simplification of Chebyshev polynomials, combines the topological structure of the graph with the node features, obtains neighbor node information through message passing, aggregates the obtained information, updates the aggregation results, and finally obtains the complex structured information in the graph.

[0095] In the above embodiment, the spatiotemporal attention convolutional network uses the mean square error (MSE) as the loss function.

[0096] The above embodiment also includes a method for evaluating the predictive ability of the spatiotemporal attention convolutional network:

[0097] The Nash efficiency coefficient NSE is used to evaluate the prediction results of the hydrological process;

[0098] NSE = 1, indicating that the model simulation values ​​are completely consistent with the observed values ​​and the model prediction effect is perfect;

[0099] NSE = 0, indicating that the model prediction effect is the same as directly using the average of the observed values ​​as the predicted value;

[0100] NSE < 0 means that the model prediction effect is worse than directly using the mean of the observed values ​​as the predicted value.

[0101] The system provided in this embodiment is used to execute the above-mentioned method embodiments. Please refer to the above-mentioned embodiments for specific processes and detailed contents, which will not be repeated here.

[0102] In one embodiment of the present invention, a computing device is provided. The computing device may be a terminal and may include: a processor, a communications interface, a memory, a display screen, and an input device. The processor, communications interface, and memory communicate with each other via a communications bus. The processor is configured to provide computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and a computer program. When executed by the processor, the computer program implements the methods described in the above embodiments. The internal memory provides an environment for the operating system and computer program in the non-volatile storage medium to run. The communications interface is configured to communicate with an external terminal via wired or wireless communication. The wireless communication may be achieved via Wi-Fi, a network management service provider, NFC (near field communication), or other technologies. The display screen may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen layer covering the display screen, or may be buttons, a trackball, or a touchpad provided on the computing device housing, or may be an external keyboard, touchpad, or mouse. The processor may invoke logic instructions stored in the memory.

[0103] In addition, the logical instructions in the above-mentioned memory can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0104] In one embodiment of the present invention, a computer program product is provided, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the methods provided by the above-mentioned method embodiments.

[0105] In one embodiment of the present invention, a non-transitory computer-readable storage medium is provided. The non-transitory computer-readable storage medium stores server instructions. The computer instructions enable a computer to execute the methods provided in the above embodiments.

[0106] The above embodiment provides a computer-readable storage medium, whose implementation principle and technical effects are similar to those of the above method embodiment, and will not be repeated here.

[0107] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0108] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0109] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A flood forecasting method, characterized in that: include: The river catchment area surrounded by the watershed line and the hydrological stations at the outlet of the catchment area are used as nodes of the topological map. The connection relationship of the topological map is constructed based on the actual flow direction of the river and the connection relationship between the tributaries, so that the organizational form of the topological map structure is adapted to the hydrological mechanism. Convert the monthly features and generate dynamic periodic timestamp vectors based on the bidirectional embedding of sine and cosine functions. Combine these vectors with the topological structure data to generate spatiotemporal coupling features. The spatiotemporal coupled feature data is fed into the spatiotemporal graph attention convolutional network training with two independent feature channels. The two independent feature channels in the spatiotemporal graph attention neural network architecture process historical data and forecast data respectively. The network supports a variable modular architecture. The meteorological data and historical flow of a certain basin during the forecast period are combined with the meteorological data through topological construction and spatiotemporal coupling to form data of the same type as the training model. The data is input into the trained spatiotemporal attention convolutional network to obtain the flood forecast results for the basin. The spatiotemporal graph attention convolutional network with dual independent feature channels includes multiple STACNN modules connected in series; Each STACNN module includes: temporal attention module, spatial attention module, spatial convolution module, temporal convolution module, residual connection module and output connection module; The input data is transmitted to the temporal attention module and the residual connection module respectively. After the temporal attention module assigns different weights to different graph structure data in the time dimension, it is multiplied with the input data and transmitted to the spatial attention module. Among them, the input data of the first-level STACNN module is the encoded graph structure data, and the input data of each subsequent STACNN module is the output data of the previous level STACNN module. The spatial attention module assigns different weights to the received data in the spatial dimension, multiplies it with the input data, and then processes it through the spatial convolution module and the temporal convolution module in sequence. It is then added to the data output by the residual connection module and output through the output connection module as the input data of the next level STACNN module. The spatial attention module is: ； in, is the output data of the spatial attention module; It is the input data of the STACNN module; is through The spatial attention score matrix after function normalization; Is the module output data; is a learnable parameter; Used to extract features in the time dimension, Used to transform the matrix dimension, mapping it from the channel dimension to the time dimension, Used to extract features in the channel dimension; is the input data after the temporal attention module.

2. The flood forecasting method according to claim 1, characterized in that: The temporal attention module is: ； in, is the time attention score matrix normalized by the function; is the output data after being processed by the temporal attention module.

3. The flood forecasting method according to claim 1, wherein: The spatial convolution module uses a graph convolution operation based on the simplification of Chebyshev polynomials to combine the topological structure of the graph with the node features, obtains neighbor node information through message passing, aggregates the obtained information, updates the aggregation results, and finally obtains the complex structured information in the graph.

4. The flood forecasting method according to claim 1, wherein: The spatiotemporal attention convolutional network uses the mean square error (MSE) as the loss function.

5. The flood forecasting method according to claim 1, wherein: Also included are methods for evaluating the predictive power of spatiotemporal attention convolutional networks: The Nash efficiency coefficient NSE is used to evaluate the prediction results of the hydrological process; NSE = 1, indicating that the model simulation value is completely consistent with the observed value, and the model prediction effect is perfect; NSE = 0, indicating that the model prediction effect is the same as directly using the average of the observed values ​​as the predicted value; NSE < 0 means that the model prediction effect is worse than directly using the mean of the observed values ​​as the predicted value.

6. A flood forecasting system for implementing the flood forecasting method according to any one of claims 1 to 5, characterized in that: include: The topological modeling module uses the river catchment area surrounded by the watershed line and the hydrological stations at the catchment outlet as the nodes of the topological map. The connection relationship of the topological map is constructed based on the actual flow direction of the river and the connection relationship between the tributaries, so that the organizational form of the topological map structure is adapted to the hydrological mechanism. The spatiotemporal coupling module converts the monthly features and generates a dynamic periodic timestamp vector based on the bidirectional embedding of the sine and cosine functions. It then combines the vector with the topological structure data to generate spatiotemporal coupling features. The spatiotemporal attention convolutional network module inputs spatiotemporal coupled feature data into the spatiotemporal graph attention convolutional network training with two independent feature channels. The two independent feature channels in the spatiotemporal graph attention neural network architecture process historical data and forecast data respectively. The network supports a variable modular architecture. The prediction module combines the meteorological data and historical flow of a certain river basin during the forecast period with the meteorological data through topological construction and spatiotemporal coupling to form data of the same type as the training model, and inputs the trained spatiotemporal attention convolutional network to obtain the flood forecast results for the river basin.

7. A computer-readable storage medium storing one or more programs, characterized in that: The one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any one of the methods of claims 1 to 5 .

8. A computing device, characterized in that include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include instructions for executing any one of the methods according to claims 1 to 5.

Images