A flood risk prediction method and system based on multi-source data fusion

Abstract

The present invention relates to a flood risk prediction method and system based on multi-source data fusion. The method comprises: obtaining current multi-source hydrological data; combining the currently obtained multi-source hydrological data with the multi-source hydrological data obtained N-1 times before, constructing a multi-source hydrological time series data set, wherein the storage form of the multi-source hydrological time series data set is a two-dimensional matrix, and the size of the multi-source hydrological time series data set is M×N, wherein M is the total number of hydrological stations; sending the multi-source hydrological time series data set into a flood prediction model for processing, and outputting a flood prediction result, wherein the flood prediction result includes a flood and no flood; the present invention combines the temporal characteristics and spatial characteristics in the multi-source hydrological time series data set to obtain a flood risk prediction method and system based on multi-source data fusion. The characteristics are analyzed, and the features of the multi-source hydrological time series data set are enhanced based on the analyzed temporal characteristics and spatial characteristics, which can improve the prediction accuracy of hydrological data and thus improve the accuracy of flood prediction; different receptive fields are set in the TCN convolution layer to analyze the temporal information of different time steps, which can effectively improve the prediction accuracy of hydrological data; and the first void convolution parameter set, the second void convolution parameter set and the third void convolution parameter set are obtained through the swarm optimization algorithm, which can improve the prediction accuracy of the flood prediction model.

Description

Technical Field

[0001] The present invention relates to the field of runoff prediction, and in particular to a flood risk prediction method and system based on multi-source data fusion. Background Art

[0002] Floods are one of the most destructive and harmful natural disasters, causing huge losses and impacts on human and socio-economic development. In the past few decades, with the intensification of climate change and urbanization, the frequency and intensity of flood events have also been on the rise, exacerbating the threat of flood risks. Therefore, timely and accurate flood risk prediction is of great significance for reducing losses, protecting people's lives and property, and the sustainable development of urban infrastructure. Summary of the invention

[0003] The present invention analyzes the temporal characteristics and spatial characteristics in a multi-source hydrological time series data set, and enhances the characteristics of the multi-source hydrological time series data set based on the analyzed temporal characteristics and spatial characteristics, thereby improving the prediction accuracy of hydrological data and further improving the accuracy of flood prediction; different receptive fields are set in the TCN convolution layer to analyze the temporal information of different time steps, which can effectively improve the prediction accuracy of hydrological data; and the first atrous convolution parameter set, the second atrous convolution parameter set and the third atrous convolution parameter set are obtained through a swarm optimization algorithm, which can improve the prediction accuracy of the flood prediction model.

[0004] A flood risk prediction method based on multi-source data fusion, including:

[0005] Access current multi-source hydrological data;

[0006] The currently acquired multi-source hydrological data are combined with the multi-source hydrological data acquired N-1 times before to construct a multi-source hydrological time series data set. The storage form of the multi-source hydrological time series data set is a two-dimensional matrix, and the size of the multi-source hydrological time series data set is M×N, where M is the total number of hydrological stations.

[0007] Send multi-source hydrological time series data sets into the flood prediction model for processing, and output flood prediction results, which include flood and no flood;

[0008] The flood prediction model is established based on the time series prediction network, and the time series prediction network adopts the TCN network. The flood prediction model includes a self-attention mechanism layer, a first feature enhancement layer, a first TCN convolution layer, a second TCN convolution layer, a third TCN convolution layer, a second feature enhancement layer, a fully connected layer and a flood prediction result output layer; wherein the self-attention mechanism layer is used to process the multi-source hydrological time series data set through the self-attention mechanism operation to construct a multi-source hydrological data self-attention weight matrix; the first feature enhancement layer is used to perform the first feature enhancement on the multi-source hydrological time series data set based on the multi-source hydrological data self-attention weight matrix to construct a first feature map; the first TCN convolution layer, the second The TCN convolution layer and the third TCN convolution layer are used to perform convolution operations on the first feature map based on the first dilated convolution parameter set, the second dilated convolution parameter set and the third dilated convolution parameter set, respectively, to construct the second feature map, the third feature map and the fourth feature map. The first dilated convolution parameter set, the second dilated convolution parameter set and the third dilated convolution parameter set are simulated by the swarm optimization algorithm; the second feature enhancement layer is used to enhance the features of the fourth feature map based on the temporal weight matrix and the spatial weight matrix corresponding to the multi-source hydrological time series data set to construct the fifth feature map; the fully connected layer and the flood prediction result output layer are used to predict floods based on the fifth feature map and output the flood prediction results.

[0009] As a preferred aspect, the multi-source hydrological time series data set is sent to the flood prediction model for processing, and the flood prediction results are output, specifically including the following results:

[0010] In the self-attention mechanism layer, the multi-source hydrological time series dataset is multiplied by the key weight matrix and the query weight matrix respectively to construct the key matrix K and the query matrix Q, and then the multi-source hydrological data self-attention weight matrix ATT=softmax(QK T / (N) 0.5 );

[0011] In the first feature enhancement layer, the multi-source hydrological time series data set is multiplied by the value weight matrix to construct the value matrix V, and then the multi-source hydrological data self-attention weight matrix ATT is multiplied by the value matrix V to construct a fusion feature map. The fusion feature map is residually connected and standardized with the first feature map to obtain the first feature map.

[0012] In the first TCN convolution layer, H first dilated convolution kernels are determined based on the first dilated convolution parameter set, the first dilated convolution parameter set includes the dilation rates corresponding to the H first dilated convolution kernels, and the first feature maps are convolved based on the H first dilated convolution kernels to obtain H first feature maps to be fused, and then the H first feature maps to be fused are subjected to a multi-scale feature fusion operation to obtain a second feature map;

[0013] In the second TCN convolution layer, H second dilated convolution kernels are determined based on the second dilated convolution parameter set, the second dilated convolution parameter set includes the dilation rates corresponding to the H second dilated convolution kernels, and the second feature maps are convolved based on the H second dilated convolution kernels to obtain H second feature maps to be fused, and then the H second feature maps to be fused are subjected to a multi-scale feature fusion operation to obtain a third feature map;

[0014] In the third TCN convolution layer, H third dilated convolution kernels are determined based on the third dilated convolution parameter set, the third dilated convolution parameter set includes dilation rates corresponding to the H third dilated convolution kernels, and the third feature maps are respectively convolved based on the H third dilated convolution kernels to obtain H third feature maps to be fused, and then the H third feature maps to be fused are subjected to a multi-scale feature fusion operation to obtain a fourth feature map;

[0015] In the second feature enhancement layer, the data vectors corresponding to each row in the multi-source hydrological time series data set are sent to N multi-layer perceptrons for processing to obtain N time feature weight vectors, and then the M time feature weight vectors are spliced ​​from top to bottom to construct a time series weight matrix; the data vectors corresponding to each column in the multi-source hydrological time series data set are sent to N multi-layer perceptrons for processing to obtain N spatial feature weight vectors, and then the N spatial feature weight vectors are spliced ​​from left to right to construct a spatial weight matrix; the result of the dot multiplication operation of the time series weight matrix and the spatial weight matrix is ​​then dot multiplied with the fourth feature map to construct the fifth feature map;

[0016] The fifth feature map is sent to the fully connected layer and the flood prediction result output layer for processing to obtain the flood prediction result.

[0017] As a preferred aspect, the training of the flood prediction model specifically includes the following steps:

[0018] A number of multi-source hydrological time series data training samples are obtained, and the multi-source hydrological time series data training samples include multi-source hydrological time series data sets in historical records. The multi-source hydrological time series data training samples are labeled by actual flood result labels, and all labeled multi-source hydrological time series data training samples are composed of a flood prediction training set. The flood prediction training set is then sent to a parameter-initialized flood prediction model for training. During this period, the actual flood result label is used as the target, and the loss value is calculated to determine whether the loss value is within a preset range. If the loss value is within the preset range, the trained flood prediction model is output; otherwise, the flood prediction model is continuously trained through the flood prediction training set.

[0019] As a preferred aspect, simulating the first dilated convolution parameter set, the second dilated convolution parameter set and the third dilated convolution parameter set by a swarm optimization algorithm specifically includes the following steps:

[0020] Construct several simulated individuals, and combine all simulated individuals into a simulated population set, and set the maximum number of iterations;

[0021] Calculate the fitness of each simulated individual, and iterate and update it based on the fitness of each simulated individual through the group optimization algorithm;

[0022] When the number of iterations reaches the maximum number of iterations, the simulated individual with the highest fitness is output as the optimal simulated individual, and the optimal simulated individual interval length H is divided to construct the first dilated convolution parameter set, the second dilated convolution parameter set and the third dilated convolution parameter set.

[0023] As a preferred aspect, the simulated individual is constructed as follows: 3H random integers are randomly generated within an expansion rate interval, and the 3H random integers are combined to form a simulated individual.

[0024] As a preferred aspect, the fitness corresponding to each simulated individual is calculated, specifically including the following contents:

[0025] The simulated individual interval length H is divided, and the first dilated convolution simulation parameter set, the second dilated convolution simulation parameter set and the third dilated convolution simulation parameter set are constructed. Then the first dilated convolution simulation parameter set, the second dilated convolution simulation parameter set and the third dilated convolution simulation parameter set are used as the first dilated convolution parameter set, the second dilated convolution parameter set and the third dilated convolution parameter set to set the first TCN convolution layer, the second TCN convolution layer and the third TCN convolution layer in the flood prediction model; several multi-source hydrological time series data test samples are obtained, and the multi-source hydrological time series data test samples are labeled with actual flood result labels. All multi-source hydrological time series data test samples are sent one by one into the flood prediction model with the first TCN convolution layer, the second TCN convolution layer and the third TCN convolution layer set for processing, and the accuracy corresponding to each multi-source hydrological time series data test sample is obtained, and the average value of the accuracy corresponding to all multi-source hydrological time series data test samples is used as the fitness corresponding to the simulated individual.

[0026] As a preferred aspect, the population optimization algorithm selects a genetic algorithm.

[0027] A flood risk prediction system based on multi-source data fusion, including:

[0028] A multi-source hydrological data set construction module is used to obtain the current multi-source hydrological data, combine the currently obtained multi-source hydrological data with the multi-source hydrological data obtained N-1 times before, and construct a multi-source hydrological time series data set;

[0029] The flood prediction module is used to send multi-source hydrological time series data sets into the flood prediction model for processing and output flood prediction results.

[0030] The present invention has the following advantages:

[0031] The present invention analyzes the temporal characteristics and spatial characteristics in a multi-source hydrological time series data set, and enhances the characteristics of the multi-source hydrological time series data set based on the analyzed temporal characteristics and spatial characteristics, thereby improving the prediction accuracy of hydrological data and further improving the accuracy of flood prediction; different receptive fields are set in the TCN convolution layer to analyze the temporal information of different time steps, which can effectively improve the prediction accuracy of hydrological data; and the first atrous convolution parameter set, the second atrous convolution parameter set and the third atrous convolution parameter set are obtained through a swarm optimization algorithm, which can improve the prediction accuracy of the flood prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a schematic diagram of the structure of a flood risk prediction system based on multi-source data fusion adopted in an embodiment of the present invention.

[0033] Figure 2 A schematic diagram of the structure of a flood prediction model used in an embodiment of the present invention;

[0034] Figure 3 The figure is a flow chart of a method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0035] In order to enable persons skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Example 1

[0036] like Figure 3 As shown in FIG. 1 , a flood risk prediction method based on multi-source data fusion includes:

[0037] Obtain the current multi-source hydrological data. The multi-source hydrological data comes from multiple hydrological stations corresponding to the river to be predicted, and the multi-source hydrological data is specifically the water level data uploaded by the hydrological station. By analyzing the water level data at different locations of the river, the spatial variation trend of the water level at different locations of the river can be obtained, thereby realizing flood prediction;

[0038] The currently acquired multi-source hydrological data are combined with the multi-source hydrological data acquired N-1 times before to construct a multi-source hydrological time series data set. The storage form of the multi-source hydrological time series data set is a two-dimensional matrix, and the size of the multi-source hydrological time series data set is M×N, where M is the total number of hydrological stations. Each row in the multi-source hydrological time series data set stores the water level data uploaded by a hydrological station arranged in time, representing the change of time series characteristics; each column in the multi-source hydrological time series data set stores the hydrological data uploaded by M hydrological stations at a time point, representing the change of spatial characteristics;

[0039] The multi-source hydrological time series data set is sent to the flood prediction model for processing, and the flood prediction results are output. The flood prediction results include flood and no flood. Flood means that the river to be tested may flood in the future, and no flood means that the river to be tested may not flood in the future. The flood prediction results can guide the advance preparation of flood control measures and reduce the losses caused by the coming of floods.

[0040] See also Figure 2The flood prediction model is established based on the time series prediction network, and the time series prediction network adopts the TCN network. The flood prediction model includes a self-attention mechanism layer, a first feature enhancement layer, a first TCN convolution layer, a second TCN convolution layer, a third TCN convolution layer, a second feature enhancement layer, a fully connected layer and a flood prediction result output layer; wherein the self-attention mechanism layer is used to process the multi-source hydrological time series data set through the self-attention mechanism operation to construct a multi-source hydrological data self-attention weight matrix. The multi-source hydrological data self-attention weight matrix can reflect the mutual influence relationship between the corresponding time series data of different hydrological stations in the multi-source hydrological time series data set, and can analyze the influence weights of the time series features and spatial features in the multi-source hydrological time series data set; the first feature enhancement layer is used to perform the first feature enhancement on the multi-source hydrological time series data set based on the multi-source hydrological data self-attention weight matrix to construct the first feature map. The first feature map can enhance the time series features and spatial features in the multi-source hydrological time series data set and improve the feature learning process in the subsequent TCN convolution process; the first TCN convolution layer, the second TCN convolution layer The first and third TCN convolution layers are used to perform convolution operations on the first feature map based on the first dilated convolution parameter set, the second dilated convolution parameter set and the third dilated convolution parameter set respectively, so as to construct the second feature map, the third feature map and the fourth feature map. The first dilated convolution parameter set, the second dilated convolution parameter set and the third dilated convolution parameter set are simulated by the swarm optimization algorithm. The second feature map, the third feature map and the fourth feature map can realize the temporal characteristics of the multi-source hydrological time series data set under the receptive fields of different scales, thereby realizing the prediction of subsequent hydrological data. The second feature enhancement layer is used to enhance the features of the fourth feature map based on the temporal weight matrix and the spatial weight matrix corresponding to the multi-source hydrological time series data set to construct the fifth feature map. The fifth feature map is the predicted multi-source hydrological data set for the next N times. The fifth feature map can further improve the utilization ability of the temporal features and spatial features in the multi-source hydrological time series data set, thereby improving the accuracy of the prediction of subsequent hydrological data. The fully connected layer and the flood prediction result output layer are used to predict floods based on the fifth feature map and output flood prediction results.

[0041] The present application analyzes the temporal characteristics and spatial characteristics in a multi-source hydrological time series data set, and enhances the characteristics of the multi-source hydrological time series data set based on the analyzed temporal characteristics and spatial characteristics, thereby improving the prediction accuracy of hydrological data and thus improving the accuracy of flood prediction; different receptive fields are set in the TCN convolution layer to analyze the temporal information of different time steps, which can effectively improve the prediction accuracy of hydrological data; and the first atrous convolution parameter set, the second atrous convolution parameter set and the third atrous convolution parameter set are obtained through a swarm optimization algorithm, which can improve the prediction accuracy of the flood prediction model.

[0042] The multi-source hydrological time series data set is sent to the flood prediction model for processing, and the flood prediction results are output, including the following results:

[0043] In the self-attention mechanism layer, the multi-source hydrological time series dataset is multiplied by the key weight matrix and the query weight matrix respectively to construct the key matrix K and the query matrix Q, and then the multi-source hydrological data self-attention weight matrix ATT=softmax(QK T / (N) 0.5 ), it should be noted that the specific operation of the self-attention mechanism layer is set with reference to the self-attention mechanism in the Transformer model. The dependency relationship of the temporal characteristics of different hydrological stations in the multi-source hydrological time series data set can be analyzed through the self-attention weight matrix of multi-source hydrological data;

[0044] In the first feature enhancement layer, the multi-source hydrological time series data set is multiplied by the value weight matrix to construct the value matrix V, and then the multi-source hydrological data self-attention weight matrix ATT is multiplied by the value matrix V to construct a fusion feature map. The fusion feature map is residually connected and standardized with the first feature map to obtain the first feature map.

[0045] In the first TCN convolution layer, H first dilated convolution kernels are determined based on the first dilated convolution parameter set. The first dilated convolution parameter set includes the expansion rates corresponding to the H first dilated convolution kernels, and H is determined by the developer. In the actual operation of building a flood prediction model, the H value can also be obtained by iterative simulation through a swarm optimization algorithm, and the size of the dilated convolution kernel is generally "1×3". Based on the H first dilated convolution kernels, convolution operations are performed on the first feature map to obtain H first feature maps to be fused, and then multi-scale feature fusion operations are performed on the H first feature maps to be fused to obtain the second feature map. The convolution method generally adopts same convolution to ensure that the size of the input data and the output data is consistent. The multi-scale feature fusion operation is generally to splice the H first feature maps to be fused according to the channel, and then perform the convolution operation;

[0046] In the second TCN convolution layer, H second dilated convolution kernels are determined based on the second dilated convolution parameter set, the second dilated convolution parameter set includes the dilation rates corresponding to the H second dilated convolution kernels, and the second feature maps are convolved based on the H second dilated convolution kernels to obtain H second feature maps to be fused, and then the H second feature maps to be fused are subjected to a multi-scale feature fusion operation to obtain a third feature map;

[0047] In the third TCN convolution layer, H third dilated convolution kernels are determined based on the third dilated convolution parameter set, the third dilated convolution parameter set includes dilation rates corresponding to the H third dilated convolution kernels, and the third feature maps are respectively convolved based on the H third dilated convolution kernels to obtain H third feature maps to be fused, and then the H third feature maps to be fused are subjected to a multi-scale feature fusion operation to obtain a fourth feature map;

[0048] In the second feature enhancement layer, the data vectors corresponding to each row in the multi-source hydrological time series data set are sent to N multi-layer perceptrons for processing to obtain N time feature weight vectors, and then the M time feature weight vectors are spliced ​​from top to bottom to construct a time series weight matrix, which can represent the different importance of time series features; the data vectors corresponding to each column in the multi-source hydrological time series data set are sent to N multi-layer perceptrons for processing to obtain N spatial feature weight vectors, and then the N spatial feature weight vectors are spliced ​​from left to right to construct a spatial weight matrix, which can represent the importance of hydrological stations at different geographical locations; the result of the dot multiplication operation of the time series weight matrix and the spatial weight matrix is ​​then dot multiplied with the fourth feature map to construct the fifth feature map;

[0049] The fifth feature map is sent to the fully connected layer and the flood prediction result output layer for processing to obtain the flood prediction result.

[0050] The training of flood prediction model includes the following steps:

[0051] A number of multi-source hydrological time series data training samples are obtained. The multi-source hydrological time series data training samples include multi-source hydrological time series data sets in historical records and are obtained by the prescription personnel according to a database that complies with legal disclosure. They can also be simulated and generated through simulation software. The multi-source hydrological time series data training samples are labeled with actual flood result labels. The actual flood result labels are flood and no flood. All labeled multi-source hydrological time series data training samples are composed of a flood prediction training set. The flood prediction training set is then sent to a parameter-initialized flood prediction model for training. During this period, the actual flood result label is used as the target, and the loss value is calculated to determine whether the loss value is within a preset range. The preset range is determined by the developer. If the loss value is within the preset range, the trained flood prediction model is output; otherwise, the flood prediction model continues to be trained through the flood prediction training set.

[0052] The first dilated convolution parameter set, the second dilated convolution parameter set and the third dilated convolution parameter set are simulated by a group optimization algorithm, specifically including the following steps:

[0053] Construct several simulated individuals, and combine all simulated individuals into a simulated population set, and set the maximum number of iterations;

[0054] The construction method of the simulated individual is as follows: 3H random integers are randomly generated within the expansion rate range, and the 3H random integers are combined into a simulated individual. The expansion rate range is set by the developer;

[0055] Calculate the fitness of each simulated individual, and iterate and update it based on the fitness of each simulated individual through a genetic algorithm, during which selection, recombination and mutation operations are performed;

[0056] Calculate the fitness of each simulated individual, including the following:

[0057] The simulated individual interval length H is divided, and the first dilated convolution simulation parameter set, the second dilated convolution simulation parameter set and the third dilated convolution simulation parameter set are constructed. Then, the first dilated convolution simulation parameter set, the second dilated convolution simulation parameter set and the third dilated convolution simulation parameter set are used as the first dilated convolution parameter set, the second dilated convolution parameter set and the third dilated convolution parameter set to set the first TCN convolution layer, the second TCN convolution layer and the third TCN convolution layer in the flood prediction model; several multi-source hydrological time series data test samples are obtained, and the multi-source hydrological time series data test samples are labeled with actual flood result labels. All multi-source hydrological time series data test samples are sent one by one to the flood prediction model with the first TCN convolution layer, the second TCN convolution layer and the third TCN convolution layer set for processing, and the accuracy corresponding to each multi-source hydrological time series data test sample is obtained, and the average value of the accuracy corresponding to all multi-source hydrological time series data test samples is used as the fitness corresponding to the simulated individual;

[0058] When the number of iterations reaches the maximum number of iterations, the simulated individual with the highest fitness is output as the optimal simulated individual, and the optimal simulated individual interval length H is divided to construct the first dilated convolution parameter set, the second dilated convolution parameter set and the third dilated convolution parameter set. Example 2

[0059] A flood risk prediction system based on multi-source data fusion, such as Figure 1 As shown, including:

[0060] The multi-source hydrological data set construction module is used to obtain the current multi-source hydrological data. The multi-source hydrological data comes from multiple hydrological stations corresponding to the river to be predicted, and the multi-source hydrological data is specifically the water level data uploaded by the hydrological station. By analyzing the water level data at different positions of the river, the spatial variation trend of the water level at different positions of the river can be obtained, thereby realizing the prediction of floods; the multi-source hydrological data currently obtained is combined with the multi-source hydrological data obtained N-1 times before to construct a multi-source hydrological time series data set. The storage form of the multi-source hydrological time series data set is a two-dimensional matrix, and the size of the multi-source hydrological time series data set is M×N, where M is the total number of hydrological stations. Each row in the multi-source hydrological time series data set stores the water level data uploaded by a hydrological station arranged in time, representing the change of time series characteristics; each column in the multi-source hydrological time series data set stores the hydrological data uploaded by M hydrological stations at a time point, representing the change of spatial characteristics;

[0061] The flood prediction module is used to send multi-source hydrological time series data sets into the flood prediction model for processing and output flood prediction results. The flood prediction results include flood and no flood. Flood means that there is a possibility of flood in the river to be tested in the future, and no flood means that there is no possibility of flood in the river to be tested in the future. The flood prediction results can guide the advance preparation of flood control measures and reduce the losses caused by floods.

[0062] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all such improvements and changes should fall within the scope of protection of the appended claims of the present invention. Parts not described in detail in this specification belong to the prior art known to those skilled in the art.

Claims

1. A flood risk prediction method based on multi-source data fusion, characterized in that: include: Access current multi-source hydrological data; The currently acquired multi-source hydrological data are combined with the multi-source hydrological data acquired N-1 times before to construct a multi-source hydrological time series data set. The storage form of the multi-source hydrological time series data set is a two-dimensional matrix, and the size of the multi-source hydrological time series data set is M×N, where M is the total number of hydrological stations. Send multi-source hydrological time series data sets into the flood prediction model for processing, and output flood prediction results, which include flood and no flood; The flood prediction model is established based on the time series prediction network, and the time series prediction network adopts the TCN network. The flood prediction model includes a self-attention mechanism layer, a first feature enhancement layer, a first TCN convolution layer, a second TCN convolution layer, a third TCN convolution layer, a second feature enhancement layer, a fully connected layer and a flood prediction result output layer; The self-attention mechanism layer is used to process the multi-source hydrological time series dataset through the self-attention mechanism operation to construct the multi-source hydrological data self-attention weight matrix; The first feature enhancement layer is used to perform the first feature enhancement on the multi-source hydrological time series dataset based on the multi-source hydrological data self-attention weight matrix to construct the first feature map; The second feature enhancement layer is used to enhance the features of the fourth feature map based on the time series weight matrix and the spatial weight matrix corresponding to the multi-source hydrological time series data set to construct the fifth feature map; The fully connected layer and the flood prediction result output layer are used to predict the flood based on the fifth feature map and output the flood prediction result; The multi-source hydrological time series data set is sent to the flood prediction model for processing, and the flood prediction results are output, including the following results: In the self-attention mechanism layer, the multi-source hydrological time series dataset is multiplied by the key weight matrix and the query weight matrix respectively to construct the key matrix K and the query matrix Q, and then the multi-source hydrological data self-attention weight matrix ATT=softmax(QK T / (N) 0.5 ); In the first feature enhancement layer, the multi-source hydrological time series data set is multiplied by the value weight matrix to construct the value matrix V, and then the multi-source hydrological data self-attention weight matrix ATT is multiplied by the value matrix V to construct a fusion feature map. The fusion feature map is residually connected and standardized with the first feature map to obtain the first feature map. In the first TCN convolution layer, H first dilated convolution kernels are determined based on the first dilated convolution parameter set, the first dilated convolution parameter set includes the dilation rates corresponding to the H first dilated convolution kernels, and the first feature maps are convolved based on the H first dilated convolution kernels to obtain H first feature maps to be fused, and then the H first feature maps to be fused are subjected to a multi-scale feature fusion operation to obtain a second feature map; In the second TCN convolution layer, H second dilated convolution kernels are determined based on the second dilated convolution parameter set, the second dilated convolution parameter set includes the dilation rates corresponding to the H second dilated convolution kernels, and the second feature maps are convolved based on the H second dilated convolution kernels to obtain H second feature maps to be fused, and then the H second feature maps to be fused are subjected to a multi-scale feature fusion operation to obtain a third feature map; In the third TCN convolution layer, H third dilated convolution kernels are determined based on the third dilated convolution parameter set, the third dilated convolution parameter set includes dilation rates corresponding to the H third dilated convolution kernels, and the third feature maps are respectively convolved based on the H third dilated convolution kernels to obtain H third feature maps to be fused, and then the H third feature maps to be fused are subjected to a multi-scale feature fusion operation to obtain a fourth feature map; In the second feature enhancement layer, the data vector corresponding to each row in the multi-source hydrological time series data set is sent to N multi-layer perceptrons for processing to obtain N time feature weight vectors, and then the M time feature weight vectors are spliced ​​from top to bottom to construct a time series weight matrix; The data vectors corresponding to each column in the multi-source hydrological time series data set are sent to N multi-layer perceptrons for processing to obtain N spatial feature weight vectors, and then the N spatial feature weight vectors are spliced ​​from left to right to construct a spatial weight matrix; The result of the dot multiplication operation of the temporal weight matrix and the spatial weight matrix is ​​then multiplied by the fourth feature map to construct a fifth feature map; The fifth feature map is sent to the fully connected layer and the flood prediction result output layer for processing to obtain the flood prediction result; The training of flood prediction model includes the following steps: Acquire several multi-source hydrological time series data training samples, which include multi-source hydrological time series data sets in historical records, annotate the multi-source hydrological time series data training samples with actual flood result labels, form a flood prediction training set with all annotated multi-source hydrological time series data training samples, and then send the flood prediction training set to a parameter-initialized flood prediction model for training, during which the actual flood result label is used as a target, and the loss value is calculated to determine whether the loss value is within a preset range. If the loss value is within the preset range, output the trained flood prediction model; otherwise, continue to train the flood prediction model with the flood prediction training set; The first dilated convolution parameter set, the second dilated convolution parameter set and the third dilated convolution parameter set are simulated by a group optimization algorithm, specifically including the following steps: Construct several simulated individuals, and combine all simulated individuals into a simulated population set, and set the maximum number of iterations; Calculate the fitness of each simulated individual, and iterate and update it based on the fitness of each simulated individual through the group optimization algorithm; When the number of iterations reaches the maximum number of iterations, the simulated individual with the highest fitness is output as the optimal simulated individual, and the optimal simulated individual interval length H is divided to construct the first dilated convolution parameter set, the second dilated convolution parameter set and the third dilated convolution parameter set; Calculate the fitness of each simulated individual, including the following: The simulated individual interval length H is divided, and the first dilated convolution simulation parameter set, the second dilated convolution simulation parameter set and the third dilated convolution simulation parameter set are constructed. Then the first dilated convolution simulation parameter set, the second dilated convolution simulation parameter set and the third dilated convolution simulation parameter set are used as the first dilated convolution parameter set, the second dilated convolution parameter set and the third dilated convolution parameter set to set the first TCN convolution layer, the second TCN convolution layer and the third TCN convolution layer in the flood prediction model; several multi-source hydrological time series data test samples are obtained, and the actual flood results are used to mark the The multi-source hydrological time series data test samples are labeled, and all the multi-source hydrological time series data test samples are sent one by one into the flood prediction model with the first TCN convolution layer, the second TCN convolution layer, and the third TCN convolution layer for processing, and the accuracy corresponding to each multi-source hydrological time series data test sample is obtained, and the average value of the accuracy corresponding to all multi-source hydrological time series data test samples is used as the fitness corresponding to the simulated individual; the construction method of the simulated individual is as follows: 3H random integers are randomly generated within the expansion rate range, and the 3H random integers are composed of simulated individuals.

2. A flood risk prediction method based on multi-source data fusion according to claim 1, characterized in that: The genetic algorithm is selected as the population optimization algorithm.

3. A flood risk prediction system based on multi-source data fusion, characterized in that: The flood risk prediction system based on multi-source data fusion applies the flood risk prediction method based on multi-source data fusion described in any one of claims 1 to 2, including: A multi-source hydrological data set construction module is used to obtain the current multi-source hydrological data, combine the currently obtained multi-source hydrological data with the multi-source hydrological data obtained N-1 times before, and construct a multi-source hydrological time series data set; The flood prediction module is used to send multi-source hydrological time series data sets into the flood prediction model for processing and output flood prediction results.

Images