Flood type prediction method based on graph convolution Transform

Abstract

The invention discloses a graph convolution Transform-based flood type prediction method, which comprises the following steps of: extracting a rainstorm event and a flood event according to flow and rainfall data observed by a hydrometric station and a rainfall station in a research area, and respectively constructing a rainstorm graph and a flood graph according to the rainstorm event and the flood event; extracting full-map comprehensive features of the flood map by using a map convolutional neural network considering an attention mechanism, determining an optimal clustering number by using an elbow rule, and then performing flood classification by using a K-means clustering method; extracting dynamic features of the rainstorm graph by using a graph convolutional neural network considering an attention mechanism, and obtaining rainstorm spatial-temporal features through Transform; and dividing all rainstorm spatial-temporal characteristics into a training set and a verification set, and training a flood type prediction model and performing flood type prediction by taking a flood type as a supervision signal. The method comprehensively considers the spatial and temporal characteristics of the rainstorm flood, and improves the prediction accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of flood classification, and specifically to a flood type prediction method based on graph convolutional Transformer. Background Technique

[0002] The flood process is affected by multiple factors such as climate, terrain, and human activities, showing significant spatio-temporal differences. The flood characteristics (such as rising and falling speeds, flood peak shapes, durations, etc.) in different basins and different seasons often vary greatly. This diversity makes it difficult for traditional analysis methods to systematically depict the unique laws of various floods, resulting in low utilization rate of historical flood information and insufficient exploration of laws.

[0003] To solve this problem, flood classification methods have emerged. Based on the similarity of rainstorm flood characteristics, historical floods are divided into several categories. By analyzing the formation mechanisms and evolution laws of various floods, a classification system is established. On this basis, by real-time monitoring indicators such as rainfall and soil moisture, combined with machine learning algorithms, rapid identification and prediction of new flood types are realized. Flood classification and flood type prediction methods not only reveal the internal laws of floods, but also provide a scientific basis for flood evolution simulation and risk assessment, and have important application values in the fields of water resources management and disaster prevention and mitigation.

[0004] Traditional flood classification methods are limited to the classification of measured floods at a single hydrological station. The flood characteristics of a single hydrological station cannot represent the situations of other regions in the entire basin, and cannot show the overall picture of floods in the entire basin. Moreover, the measured data of a single hydrological station are easily interfered by local special factors. Summary of the Invention

[0005] The purpose of the present invention is to provide a flood type prediction method based on graph convolutional Transformer to judge the overall flood occurrence situation and flood types from the perspective of the entire basin.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: A flood type prediction method based on graph convolutional Transformer, the steps of which include: Extract rainstorm events and flood events according to the flow and precipitation data observed by hydrological stations and rain gauges in the study area; specifically, consider that a rainstorm event occurs when the total precipitation observed by a single rain gauge in 24 hours is greater than 50 mm or the total precipitation observed in 12 hours is greater than 30 mm; the measured flow of a single hydrological station is greater than 100 m 3 / s is regarded as a flood event. Among them, analyze the occurrence time and correlation relationship between the observed rainstorm events at the rainfall stations and the measured flood events at the hydrological stations. If a corresponding flood event occurs after a rainstorm event, it is regarded as a valid rainstorm event; if no corresponding flood event occurs after a rainstorm event, it is not included in the statistics of rainstorm events.

[0007] Construct a rainstorm map and a flood map based on the rainstorm events and flood events respectively; the node characteristics of the flood map are the various statistical characteristic index values of the measured flood at the corresponding hydrological station in this flood event. One flood map corresponds to one flood event, and directly extract the multi-dimensional characteristics of this flood map to represent the flood spatio-temporal characteristics; The node characteristics of the rainstorm map are the period precipitation (dynamic characteristics) and land use and slope (static characteristics) of the grid where the node is located. One rainstorm map corresponds to one time step.

[0008] For example, a rainstorm event starts at 8 am and lasts until 11 am. If the time step is 1h, then there are 3 time steps, corresponding to 3 rainstorm maps. Extract the characteristics of these 3 rainstorm maps to obtain 3 sets of multi-dimensional characteristics, and form a characteristic sequence in chronological order to represent the rainstorm spatio-temporal characteristics.

[0009] Use a graph convolutional neural network considering the attention mechanism to extract the comprehensive global characteristics of the flood map, use the elbow method to determine the optimal number of clusters for multiple comprehensive global characteristics, and use the K-means clustering method to classify the floods to obtain the corresponding flood types; Use a graph convolutional neural network considering the attention mechanism to extract the dynamic characteristics of the rainstorm map, and capture the time dependence relationship of the dynamic characteristics of the rainstorm map through a Transformer. After aggregation, obtain the rainstorm spatio-temporal characteristics; Divide all rainstorm spatio-temporal characteristics into a training set and a validation set. Using the flood type as a supervision signal, learn the correlation between the rainstorm spatio-temporal characteristics and the flood type, and train a flood type prediction model; Use the trained flood type prediction model to predict the flood type.

[0010] According to the above technical solution, the steps for constructing the rainstorm map include: Discretize the basin of the study area into 1km×1km grids. Taking the basin grids as nodes, the precipitation process of the grid rainstorm event as the node dynamic characteristics, and the land use degree and slope of the grid as the static characteristics of the rainstorm map nodes; Use the inverse distance weighting method to interpolate the precipitation at the rainfall stations to each grid to obtain the precipitation characteristic values of the nodes. The interpolation formula is: ; In the formula, is the predicted value of the point to be interpolated , For the known sample point i, is the number of the nearest neighbor points, denotes the observed value at

[0011] The interpolation weight is calculated as follows: ; wherein, is the Euclidean distance between the point to be interpolated and the known sample point, is the power parameter.

[0012] Based on the flow direction relationship between the grids, a directed edge is established between the nodes of the rainstorm map to construct the rainstorm map.

[0013] Specifically, the D8 flow direction method is used to determine the flow direction of the grid. The output value range of the D8 direction type is an integer grid between 1 and 255. For example, if the steepest descent direction is on the left side of the currently processed grid, the flow direction value of this processed grid is 16.

[0014] According to the above technical solution, the steps for constructing the flood map include: Taking the hydrological station as a node, and the characteristic values of the measured flood events at the hydrological station as the characteristics of the flood map nodes; the characteristic values of the measured flood events include peak flow, flood duration, total flood volume, maximum 1-hour flood volume, and maximum 3-hour flood volume; Based on the water system topological relationship between the hydrological stations, a directed edge is established between the nodes of the flood map to construct the flood map.

[0015] Among them, the peak flow : ; The flood duration : ; The total flood volume : ; The maximum 1-hour flood volume : ; The maximum 3-hour flood volume : ; Among them, represents the flow rate sequence of this flood event, represents the start time of the flood event, represents the end time of the flood event, represents the flow rate at the i-th moment, represents the flow rate at the (i + 1)-th moment, represents the flow rate observation interval.

[0016] According to the above technical solution, the flood type determination step includes: Collect the characteristic data of all flood events and all nodes in the flood map, independently calculate the mean and standard deviation of each characteristic dimension of all nodes, and then perform Z-score standardization processing respectively; taking the peak flood flow of node characteristics as an example, it is to normalize the peak flood flow of all nodes; Based on the standardized flood events and the characteristic data of nodes, use a graph convolutional neural network considering the attention mechanism to capture the spatial relationship between flood events, and use attention weights to weighted-aggregate neighbor features to obtain multi-dimensional features of each node; Input the multi-dimensional features of all nodes into the global average pooling to obtain the comprehensive feature of the entire flood map of a single flood map; ; In the formula, is the number of flood map nodes, is the -th node at the -th dimension of the eigenvalue, represents the -th dimension of the global graph-level feature of the flood map.

[0017] Determine the optimal number of clusters according to the elbow method. The core idea is to observe the change trend of the clustering error SSE with the increase of the K value and find the "inflection point" (i.e., the "elbow"). The calculation formula is as follows: ; Among them, represents the -th cluster, represents the centroid of the -th cluster, represents the multi-dimensional global graph-level feature of a single flood map.

[0018] Based on the optimal number of clusters determined by the elbow method, use the K-means clustering method to cluster the comprehensive features of the entire flood maps of multiple flood maps, and use the clustering result as the flood type classification result to determine the flood type.

[0019] Among them, in the graph convolutional neural network considering the attention mechanism, the feature representation of its nodes is similar to that of the nodes in the ordinary graph neural network. Both use the embedding method to vectorize the feature representation of nodes, and its initial input is also the feature combination of each node. The initial input is expressed by the formula: ; where is the input feature of node z, represents the feature combination, represents the input feature of node 1, represents the input feature of node 2, represents node 's input feature, is the dimension of the node features in the graph structure.

[0020] After the features of the nodes in the input graph, operations are performed in the graph convolutional neural network considering the attention mechanism. For node and neighbor , first calculate the attention score : ; In the formula, represents the input feature of node , represents the input feature of neighbor , is a learnable weight matrix (feature transformation), is a learnable attention vector, is vector concatenation.

[0021] Based on the calculated attention scores of node and neighbor , normalize the attention weights and weighted aggregate the neighbor features to obtain the multi-dimensional features of each node: ; ; In the formula, represents the attention weight of node , represents the output feature of the graph convolutional neural network of node , is the activation function (such as ReLU), represents the set of all neighbors j of node , .

[0022] According to the above technical solution, the steps for obtaining the comprehensive feature of the entire rainstorm map include: Divide a rainstorm event into rainstorm maps of multiple time steps according to the time step. The graph data structure of each rainstorm map is the same. The static features of the node features include the slope and land use degree and remain unchanged, and the dynamic feature is the precipitation of the corresponding time step; Collect the feature data of all rainstorm events and all nodes in the rainstorm map, independently calculate the mean and standard deviation of each feature dimension of all nodes and then perform standardization processing; Capture the spatial distribution characteristics and water volume exchange law of the basin precipitation of the rainstorm maps at multiple time steps after standardization by using a graph convolutional neural network considering the attention mechanism, and obtain the dynamic characteristics of the rainstorm maps for each node; Use Transformer to process the dynamic characteristics of the rainstorm maps for each node in the rainstorm map, and obtain the multi-dimensional characteristics containing spatio-temporal fusion information for each node in the rainstorm map; Based on the multi-dimensional characteristics of all nodes in the rainstorm map, use the attention graph pooling method to obtain the rainstorm spatio-temporal characteristics of a single rainstorm event at multiple time steps.

[0023] Among them, in the attention graph pooling, the importance of nodes is learned through a fully connected layer network and Tanh activation, and the mathematical expression is as follows: ; Among them, and are the parameters of the first fully connected layer, is the hyperbolic tangent activation function, and are the parameters of the second fully connected layer, represents the original attention score of node i.

[0024] Then, the attention scores of all nodes are normalized through the softmax function: ; Finally, the rainstorm spatio-temporal characteristics are obtained: ; In the formula, represents the original attention score of node i, represents the final attention score of node i, represents the output feature of the graph convolutional neural network of node i, represents the number of nodes in the rainstorm map.

[0025] According to the above technical solution, use a fixed random seed to ensure that the division results of the training set and the validation set are consistent each time, and the distribution of flood types in the training set and the validation set is the same.

[0026] According to the above technical solution, using the flood type label as the supervision signal, use the Adam optimizer to perform multi-layer perceptron classification training, calculate the error between the prediction result of the flood type prediction model and the true flood type during each training, and adjust the parameters of the flood type prediction model according to the error; Record the loss values and evaluation metrics of the training set and the validation set during the entire training process of the flood type prediction model in real time, and obtain the model parameters with the best performance on the validation set as the final flood type prediction model parameters.

[0027] According to the above technical solution, the loss function in the training process of the Adam optimizer is the cross-entropy loss function, and the cross-entropy loss function Calculation formula: ; In the formula, is the number of categories, is the true label of the -th category, is the predicted probability of the model for the -th category.

[0028] According to the above technical solution, the evaluation index : ; Among them, is the total number of samples, is the predicted category of the model for the -th sample, is the predicted category of the -th sample, is the indicator function, which returns 1 when the flood type label matches, and 0 otherwise.

[0029] Compared with the prior art, the beneficial effects achieved by the present invention are: 1. The present invention makes full use of the observed data of each hydrological station and rainfall station in the basin, as well as the underlying surface characteristic data, and to a certain extent considers the characteristics of the underlying surface spatial distribution map and the spatio-temporal evolution characteristics of rainstorms and floods.

[0030] 2. The present invention uses a graph data structure to represent the rainstorm flood process, realizes the synchronous capture and deep fusion of the spatio-temporal characteristics of rainstorm floods, and can better reflect the non-linear response relationship between rainfall and runoff in the basin compared with traditional classification methods.

[0031] 3. The present invention optimizes the number of clusters through the elbow method, avoids the subjectivity of artificially setting the number of classifications, and pushes flood classification from static statistics to dynamic intelligent prediction, providing a new technical paradigm for the construction of intelligent water conservancy.

[0032] 4. After using the graph data structure to represent the rainstorm flood process, the present invention uses the flood type as a supervision signal to learn the correlation between the comprehensive characteristics of the entire rainstorm map and the flood type, and conducts flood classification. In the process of flood classification and flood type prediction, the spatio-temporal characteristics of rainstorm floods are comprehensively considered, and the overall flood occurrence situation and flood type can be judged from the perspective of the entire basin, providing a more scientific decision-making basis for basin water resources management and flood control and disaster reduction. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The accompanying drawings are used to provide a further understanding of the present invention and form a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the accompanying drawings: Figure 1 is a flowchart of the steps of a flood type prediction method based on graph convolutional Transformer of the present invention; Figure 2 is a graph of the optimal clustering result of the elbow method in the embodiment; Figure 3 is a comparison graph of the characteristics of different types of floods in the example. Detailed implementation manners

[0034] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0035] Taking the Yiluo River Basin, a tributary in the middle reaches of the Yellow River, as an example, the flow and precipitation data used are the flood element extraction data of 15 hydrological stations in the Yiluo River Basin and the precipitation extraction data of 135 rain gauge stations. The time range is from 2020 to 2024. A flood type prediction method based on graph convolutional Transformer proposed by the present invention is used to predict the flood type. The specific steps ( Figure 1 ) include: S1. Extract rainstorm events and flood events according to the flow and precipitation data observed by the hydrological stations and rain gauge stations in the study area. Specifically, when the measured flow of a single hydrological station is greater than 100 m 3 / s, it is regarded as a flood event occurring. When the total precipitation observed by a single rain gauge station in 24 hours is greater than 50 mm or the total precipitation observed in 12 hours is greater than 30 mm, it is regarded as a rainstorm event occurring. And analyze the occurrence time and correlation relationship between the rainstorm events observed by the rain gauge stations and the flood events measured by the hydrological stations. For example, if a corresponding flood event occurs after a rainstorm event, it is regarded as a valid rainstorm event; if no corresponding flood event occurs after a rainstorm event, it is not included in the statistics of rainstorm events. It is statistically found that there are 26 rainstorm flood events in the Yiluo River Basin, a tributary in the middle reaches of the Yellow River.

[0036] S2. Construct a rainstorm graph and a flood graph according to the rainstorm events and flood events respectively. Twenty-six flood graphs are constructed according to 26 rainstorm flood events.

[0037] S2a. The steps for constructing the rainstorm map are specifically as follows: The basin is discretized into 1 km × 1 km grids. Taking the basin grids as nodes, the precipitation of the rain gauge stations is interpolated to each grid using IDW to obtain the precipitation characteristic values of the nodes. The precipitation process of the grid rainstorm event is the dynamic characteristic of the nodes, and the grid land use degree and slope are the static characteristics of the rainstorm map nodes.

[0038] The D8 flow direction method is used to determine the flow direction of the grids. The output value range of the D8 direction type is an integer grid between 1 and 255. For example, if the steepest descent direction is on the left side of the currently processed grid, the flow direction value of this processed grid is 16.

[0039] Based on the flow direction relationship between the grids, a directed edge is established between the rainstorm map nodes to construct the rainstorm map.

[0040] Among them, the interpolation formula for interpolating the precipitation of the rain gauge stations to each grid using IDW is: ; In the formula, is the predicted value of the point to be interpolated, is the value of the known sample point, is the number of the nearest neighbor points, is the interpolation weight, and its calculation formula is as follows: ; In the formula, is the Euclidean distance between the point to be interpolated and the known sample point, is the power parameter. Among them, the number of the nearest neighbor points and the power parameter take the values of 10 and 2 respectively.

[0041] The land use degree is represented by the land use comprehensive index, and its calculation formula is as follows: ; Among them, is the comprehensive index of the land use degree in the study area; is the land use classification index of the i-th level; is the proportion of the land area of the i-th level of land use. According to the degree of land development and utilization by human activities, the land use types are divided into 4 levels, namely the unused land level (classification index is 1), the forest, grass, and water use land level (classification index is 2), the agricultural land level (classification index is 3), and the urban settlement land level (classification index is 4).

[0042] The slope can be calculated using ArcGIS based on DEM, and the calculation method is the third-order finite difference (3×3 window) method.

[0043] S2b. The steps for constructing the flood map specifically include: taking hydrological stations as nodes, and using the characteristic values of the measured flood events at the hydrological stations as the node characteristics of the flood map; the characteristic values of the measured flood events include peak flow, flood duration, flood volume, maximum 1-hour flood volume, and maximum 3-hour flood volume. Based on the river system topological relationship between hydrological stations, establish directed edges between the nodes of the flood map to construct the flood map.

[0044] S3. For the 26 flood maps, use the graph convolutional neural network (GAT) considering the attention mechanism to extract the comprehensive global features of the flood maps, determine the optimal number of clusters for the 26 comprehensive global features using the elbow method, and use the K-means clustering method for flood classification to obtain the corresponding flood types. The specific steps include: S3a. Collect the feature data of all flood events and all nodes in the flood map, independently calculate the mean and standard deviation of each feature column, and then perform Z-score standardization processing respectively. S3b. Based on the standardized feature data of flood events and nodes, use the graph convolutional neural network considering the attention mechanism to capture the spatial relationship between flood events, and use attention weights to weighted-aggregate neighbor features to obtain the multi-dimensional features of each node.

[0045] Specifically: The feature representation of the nodes in the flood map in the graph convolutional neural network considering the attention mechanism is similar to that of the nodes in the ordinary graph neural network. Both use the embedding method to vectorize the feature representation of the nodes, and their initial input is also the feature combination of each node.

[0046] After inputting the features of the nodes in the flood map, perform operations inside the graph convolutional neural network considering the attention mechanism. For node and neighbor , first calculate the attention score; According to the attention scores of node and its neighbors, normalize the attention weights and weighted-aggregate the neighbor features to obtain the multi-dimensional features of each node.

[0047] S3c. Input the multi-dimensional features of all nodes into the global average pool to obtain the comprehensive global feature of a single flood map; S3d. Determine the optimal number of clusters according to the elbow method. Its core idea is to observe the change trend of the clustering error SSE with the increase of the K value and find the "inflection point" (i.e., the "elbow"). The calculation formula is as follows: ; where represents the th cluster. represents the centroid of the nth cluster, indicating the multi-dimensional global map-level features of a single flood map.

[0048] S3e. Determine the optimal number of clusters based on the elbow method. Use the K-means clustering method to cluster the comprehensive full-map features of multiple flood maps, and take the clustering result as the flood type classification result to determine the flood type. The optimal number of clusters determined by the elbow method for 26 comprehensive full-map features is 3. The comparison of the flood feature distributions of three different types of floods is shown in Figure 3 , and it can be seen that the characteristics of the three types of floods in the Yiluo River are as follows: (1) Type 0 flood: Short-duration heavy rainfall flood, with a high peak but short duration, caused by local summer rainstorms.

[0049] (2) Type 1 flood: Higher peak flow and longer duration, caused by persistent precipitation (heavy rain superimposed on the previous soil moisture condition) or extreme precipitation events.

[0050] (3) Type 2 flood: Long-term continuous flood, with a large total flood volume and long duration, but lower peak and short-term water volume, caused by continuous autumn rain.

[0051] S4. Use a graph convolutional neural network considering the attention mechanism to extract the dynamic features of the rainstorm map, and capture the temporal dependence relationship of the dynamic features of the rainstorm map through a Transformer, and aggregate to obtain the rainstorm spatio-temporal features. Specifically, Divide a rainstorm event into rainstorm maps of multiple time steps according to the time step. The graph data structure of each rainstorm map is the same. The static features of the node features include the slope and land use degree and remain unchanged, and the dynamic feature is the precipitation of the corresponding time step; Use a graph convolutional neural network considering the attention mechanism to capture the spatial distribution characteristics and water volume exchange law of the basin precipitation of the rainstorm maps of multiple time steps, and obtain the dynamic features of the rainstorm map of each node; Use a Transformer to process the dynamic features of the rainstorm map of each node in the rainstorm map, and obtain the multi-dimensional features containing spatio-temporal fusion information of each node in the rainstorm map; Based on the multi-dimensional features of all nodes in the rainstorm map, use the attention graph pooling method to obtain the rainstorm spatio-temporal features of multiple time steps of a single rainstorm event. Among them, in the attention graph pooling, the importance of nodes is learned through a two-layer network and Tanh activation, and the attention scores of all nodes are normalized through the softmax function, and finally the rainstorm spatio-temporal features are obtained.

[0052] S5. Divide all the spatio-temporal characteristics of heavy rain into a training set and a validation set according to the flood type ratio. The division ratio of the training set and the validation set is 80% and 20%. Using the flood type as the supervision signal, use the Adam optimizer to perform multi-layer perceptron classification training to learn the association between the spatio-temporal characteristics of heavy rain and the flood type, and train the flood type prediction model. Calculate the error between the model prediction result and the real flood type during each training, and adjust the parameters of the flood type prediction model according to the error.

[0053] The loss function during the training process of the Adam optimizer is the cross-entropy loss function, the cross-entropy loss function Calculation formula: ; In the formula, is the number of categories, is the true label of the th category, is the predicted probability of the model for the th category.

[0054] Record in real time the loss values and evaluation metrics of the training set and the validation set during the entire training process of the flood type prediction model, and obtain the model parameters with the best performance on the validation set as the final flood type prediction model parameters.

[0055] Among them, the evaluation metric : ; Among them, is the total number of samples, is the predicted category of the model for the th sample, is the th sample's predicted category, is the indicator function, which returns 1 when the flood type label matches, otherwise returns 0.

[0056] S6. Use the trained flood type prediction model to predict the flood type in the Yiluo River Basin, a tributary in the middle reaches of the Yellow River.

[0057] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the term "including", "comprising" or any other variation thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device.

[0058] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions recorded in the foregoing embodiments or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A flood type prediction method based on graph convolutional Transformer, characterized in that The steps include: Extracting rainstorm events and flood events based on the flow and precipitation data observed by hydrological stations and rainfall stations in the study area; Constructing a rainstorm map and a flood map according to the rainstorm events and flood events respectively; Using a graph convolutional neural network considering the attention mechanism to extract the comprehensive global features of the flood map, determining the optimal number of clusters for multiple comprehensive global features using the elbow method, and classifying floods using the K-means clustering method to obtain the corresponding flood types; Using a graph convolutional neural network considering the attention mechanism to extract the dynamic features of the rainstorm map, and capturing the temporal dependence of the dynamic features of the rainstorm map through a Transformer, and aggregating to obtain the spatio-temporal features of the rainstorm; Dividing all spatio-temporal features of the rainstorm into a training set and a validation set, using the flood type as a supervision signal to learn the association between the spatio-temporal features of the rainstorm and the flood type, and training a flood type prediction model; Using the trained flood type prediction model to predict the flood type.

2. The flood type prediction method based on graph convolutional Transformer according to claim 1, characterized in that, The steps for constructing the rainstorm map include: Discretizing the basin of the study area into grids, taking the basin grids as nodes, the precipitation process of the grid rainstorm event as the dynamic feature of the rainstorm map node, and the land use degree and slope of the grid as the static features of the rainstorm map node; Based on the flow direction relationship between the grids, establishing directed edges between the rainstorm map nodes to construct the rainstorm map.

3. A flood type prediction method based on graph convolutional Transformer according to claim 1, characterized in that, The steps for constructing the flood map include: Taking hydrological stations as nodes, and the characteristic values of the measured flood events at the hydrological stations as the node features of the flood map; the characteristic values of the measured flood events include peak flow, flood duration, total flood volume, maximum 1-hour flood volume, and maximum 3-hour flood volume; Based on the water system topological relationship between the hydrological stations, establishing directed edges between the flood map nodes to construct the flood map.

4. A flood type prediction method based on graph convolutional Transformer according to claim 1, characterized in that, The steps for determining the flood type include: Collecting the feature data of all flood events and all nodes in the flood map, independently calculating the mean and standard deviation of each feature dimension of all nodes and then performing standardization processing; Based on the standardized feature data of the flood events and nodes, using a graph convolutional neural network considering the attention mechanism to capture the spatial relationship between the flood events, and weighted aggregating the neighbor features with attention weights to obtain the multi-dimensional features of each node; Inputting the multi-dimensional features of all nodes into the global average pool to obtain the comprehensive global feature of a single flood map; Determining the optimal number of clusters according to the elbow method; Based on the optimal number of clusters determined by the elbow method, using the K-means clustering method to cluster the comprehensive global features of multiple flood maps, and taking the clustering result as the flood type classification result to determine the flood type.

5. A flood type prediction method based on graph convolutional Transformer according to claim 1, characterized in that, The steps for obtaining the comprehensive global feature of the rainstorm map include: Dividing a rainstorm event into multiple time-step rainstorm maps according to the time step, the graph data structure of each rainstorm map is the same, the static features of the node features include the slope and land use degree and remain unchanged, and the dynamic feature is the precipitation of the corresponding time step; Collecting the feature data of all rainstorm events and all nodes in the rainstorm map, independently calculating the mean and standard deviation of each feature dimension of all nodes and then performing standardization processing; Capture the spatial distribution characteristics and water volume exchange laws of the basin precipitation of the rainstorm maps at multiple time steps after normalization by using a graph convolutional neural network considering the attention mechanism, and obtain the dynamic characteristics of the rainstorm maps for each node; Process the dynamic characteristics of the rainstorm maps for each node in the rainstorm maps by using a Transformer to obtain the multi-dimensional characteristics containing spatio-temporal fusion information for each node in the rainstorm maps; Based on the multi-dimensional characteristics of all nodes in the rainstorm maps, use the attention graph pooling method to obtain the rainstorm spatio-temporal characteristics at multiple time steps for a single rainstorm event.

6. The flood type prediction method based on graph convolutional Transformer according to claim 1, wherein Use a fixed random seed to ensure that the division results of the training set and the validation set are consistent each time, and the distributions of flood types in the training set and the validation set are the same.

7. A flood type prediction method based on graph convolutional Transformer according to claim 1, characterized in that, Using the flood type label as the supervision signal, use the Adam optimizer to perform multi-layer perceptron classification training. Calculate the error between the prediction result of the flood type prediction model and the true flood type during each training, and adjust the parameters of the flood type prediction model according to the error; Record the loss values and evaluation metrics of the training set and the validation set during the entire training process of the flood type prediction model in real time, and obtain the model parameters with the best performance on the validation set as the final flood type prediction model parameters.

8. A flood type prediction method based on graph convolutional Transformer according to claim 7, characterized in that, The loss function in the training process of the Adam optimizer is the cross-entropy loss function, and the cross-entropy loss function Calculation formula: ； In the formula, is the number of categories, is the true label of the -th category, is the predicted probability of the model for the -th category.

9. A flood type prediction method based on graph convolutional Transformer according to claim 7, characterized in that, The evaluation metrics : ； Among them, is the total number of samples, is the predicted category of the model for the th sample, is the predicted category of the th sample, is the indicator function that returns 1 when the flood type label matches, and 0 otherwise.

Images