Typhoon storm surge beach overtopping intelligent forecasting method based on Vision Transform neural network

Abstract

The invention discloses a typhoon storm surge beach overtopping intelligent forecasting method based on a Vision Transform neural network. The method comprises the following steps: firstly, acquiring continuous storm surge water level data and wind field data at a forecasting moment; preprocessing the acquired continuous storm surge water level data and the wind field data at the forecast moment, performing extreme value normalization on the preprocessed data, scaling the data to 0-1, and dividing a training set and a test set; based on the training set, the normalized numerical value is input into a Vision transfer neural network for training, and a typhoon storm surge beach overtopping intelligent forecasting model is obtained; and inputting the test set into the typhoon storm surge beach overtopping intelligent forecasting model to adjust and optimize the typhoon storm surge beach overtopping to obtain a final typhoon storm surge beach overtopping intelligent forecasting model, and performing typhoon storm surge beach overtopping forecasting by using the typhoon storm surge beach overtopping intelligent forecasting model. According to the method, the two-dimensional distribution and the water increasing height of the typhoon storm surge overtopping beach can be accurately and quickly predicted.

Description

Technical Field

[0001] The present invention relates to the field of hydrological technology, and in particular to a typhoon storm surge floodplain intelligent forecasting method based on a Vision Transformer neural network. Background Art

[0002] When severe meteorological disturbances occur, the seawater tide may rise abnormally. This phenomenon is called storm surge, and in some cases it is also called "storm surge" or "storm tsunami". Whether a storm surge will become a disaster depends on whether the seawater rise caused by climate turbulence coincides with the astronomical high tide in the tidal cycle. Various climate events, including typhoons, temperate cyclones, and cold currents and strong winds, can be the cause of storm surges in specific areas. According to their causes, storm surges can be divided into two categories: storm surges in temperate regions and storm surges caused by typhoons. Usually, typhoon storm surges occur more in summer and autumn, with rapid and destructive situations. According to the data of the "China Marine Disaster Bulletin", the direct economic losses caused by typhoon storm surges rank first among various marine disasters in China. Therefore, timely and effective forecasting and warning of storm surges are extremely important and urgent in terms of national economic security and the protection of life and property of the public. Summary of the invention

[0003] The purpose of the present invention is to provide a storm surge floodplain intelligent prediction method based on Vision Transformer, which uses a deep learning method to predict a two-dimensional storm surge floodplain for a long time.

[0004] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0005] A typhoon storm surge floodplain intelligent forecasting method based on Vision Transformer neural network, characterized in that the steps of the intelligent forecasting method include:

[0006] Obtaining continuous storm surge water level data and wind field data at the forecast time; the wind field data includes the component u of the wind speed in the east-west direction and the component v of the wind speed in the north-south direction;

[0007] Preprocess the continuous storm surge water level data and wind field data at the forecast time, normalize the preprocessed data to the maximum value, and then divide them into training set and test set;

[0008] Based on the training set, the normalized values ​​are input into the Vision transformer neural network for training to obtain the typhoon storm surge floodplain intelligent forecasting model; the test set is input into the typhoon storm surge floodplain intelligent forecasting model to adjust and optimize the typhoon storm surge floodplain to obtain the final typhoon storm surge floodplain intelligent forecasting model;

[0009] The typhoon storm surge floodplain intelligent forecasting model is used to forecast the typhoon storm surge floodplain based on continuous storm surge water level data and wind field data at the forecast time.

[0010] According to the above technical solution, the preprocessed data is a combination of an array of continuous storm surge water level data and an array of wind field data at the forecast time into one sample. The preprocessed data is four-dimensional data, and the four-dimensional data is data volume*number of channels*280*190.

[0012] According to the above technical solution, the maximum value normalization is to scale the data to between 0 and 1.

[0013] According to the above technical solution, the Vision transformer neural network structure consists of an image block embedding layer, 3 encoding layers and a convolutional layer;

[0014] The output end of the image block embedding layer is connected to the input end of the encoding layer, and the output end of the encoding layer is connected to the input end of the convolutional layer; the three encoding layers are connected in series.

[0015] According to the above technical solution, the image block embedding layer is used to cut the image into patches of a fixed size of 10x10, and the dimension of the image embedding layer is set to 1024;

[0016] The encoding layer includes a multi-head attention mechanism, the number of heads in each multi-head attention mechanism is 64, the dimension of each head in the multi-head attention mechanism is set to 64, and the dimension of the middle hidden layer of the multi-layer perceptron is 1024; the multi-head attention mechanism enables the model to process data in parallel and pay attention to different information dimensions, while the subsequent multi-layer perceptron strengthens the nonlinear representation ability of the model and improves the ability to handle complex relationships;

[0017] The convolution layer uses a 1x1 convolution kernel to change the number of channels of the feature map without changing the spatial dimension.

[0018] According to the above technical solution, the prediction steps of the typhoon storm surge floodplain intelligent forecasting model include:

[0019] The image data in the training set is input into the image block embedding layer, which cuts the image data into patches of a fixed size of 10x10 (converting the image into a sequence), and then uses a trainable linear projection matrix E to map the vectorized image blocks into a potential embedding space vector with D dimensions to obtain the embedding vector of the image blocks, and then adds a corresponding D-dimensional position embedding vector generated for each image block to the embedding vector of the corresponding image block to obtain the final image patch embedding vector to preserve the spatial position information of the image block in the original image;

[0020] The image patch embedding vector is input into the encoding layer, and the data is processed in parallel by a multi-head attention mechanism with a dimension of 64, focusing on different information dimensions, and then the nonlinear representation capability is enhanced by a multi-layer perceptron with a hidden layer dimension of 1024;

[0021] After passing through three identical encoding layers, the data is input into the convolution layer and operated with a 1x1 convolution kernel to output the typhoon storm surge floodplain forecast.

[0022] Among them, the trainable linear projection matrix E is automatically generated inside the model, and the value is random. During the training process, through the back propagation algorithm, the value of the matrix E will be continuously adjusted to minimize the loss function; the vectorized image block refers to flattening the two-dimensional image block into a one-dimensional vector form.

[0023] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: the present invention proposes an intelligent storm surge floodplain prediction method based on VisionTransformer, and uses a deep learning method to predict the two-dimensional storm surge floodplain for a long time. The present invention can accurately and quickly predict the two-dimensional distribution and water surge height of the typhoon storm surge floodplain. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0025] Figure 1 It is a structural example of the present invention for predicting typhoon storm surge floodplains based on Vision transformer neural network;

[0026] Figure 2 It is a two-dimensional floodplain and water height map of typhoon storm surge predicted based on the verification data of Typhoon No. 1707. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in 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 embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] The input data used in Example 1 are the continuous storm surge water level data for the past 7 hours and the u and v wind field data at the forecast time, with a size of 280*190. A typhoon storm surge floodplain intelligent forecasting method based on a vision transformer neural network includes the following steps:

[0029] Step 1: Preprocess the two-dimensional typhoon storm surge floodplain data and the wind field data of the wind speed component u in the east-west direction and the wind speed component v in the north-south direction from 1979 to 2009 to obtain a four-dimensional data (data volume * 9 * 280 * 190), then normalize it to the maximum value and scale the data to between 0 and 1.

[0030] Step 2: Input the data processed in step 1 into the model for training. The input sample is first input into the image block embedding layer with a dimension of 1024. After the image is cut into patches of a fixed size of 10x10 in this layer, the vectorized image block is mapped to a D-dimensional latent embedding space vector using a trainable linear projection matrix E (automatically generated within the model, the value is random. During the training process, the value of the matrix E will be continuously adjusted through the back propagation algorithm to minimize the loss function) to obtain the embedding vector of the image block. Then, a corresponding D-dimensional position embedding vector generated for each image block is added to the embedding vector of the corresponding image block to obtain the final image patch embedding vector to preserve the spatial position information of the image block in the original image.

[0031] Step 3: The image patch embedding vector processed and converted in step 2 is then input into the encoding layer. After passing through a multi-head attention mechanism with a dimension of 64, the model can process data in parallel and focus on different information dimensions. The multi-layer perceptron with a hidden layer dimension of 1024 enhances the nonlinear representation capability of the model and improves its ability to handle complex relationships.

[0032] Step 4: After passing through three identical encoding layers, a 1x1 convolution kernel is input for operation to change the number of channels of the feature map without changing its spatial dimension to form the output.

[0033] Step 5: Use steps 1 to 4 to train a typhoon storm surge floodplain intelligent forecasting model, apply the verification data set to the trained typhoon storm surge floodplain intelligent forecasting model, repeat steps 2 to 4, and obtain the typhoon storm surge floodplain forecast predicted by the Vision Transformer neural network.

[0034] The 500 storm surge floodplain events in the test set were evaluated with the FVCOM data ( Figure 2 ), Figure 2 VIT in the figure represents the two-dimensional floodplain and water-increase height map of typhoon storm surge obtained by the present invention. Figure 2 FVCOM refers to the two-dimensional floodplain and water height map of typhoon storm surge obtained by using FVCOM. The results show that the MAE of the prediction results of the model of the present invention compared with the FVCOM data reaches 0.08; the model of the present invention can well reproduce the floodplain conditions simulated by FVCOM, and can make better predictions on the regional and SSH distribution of the floodplain. Compared with the complexity of the dynamic drive model, Vision Transformer is more convenient, which greatly saves the time cost of forecasting typhoon storm surge floodplains, and can accurately depict the distribution of floodplain events.

[0035] Example 2

[0036] The Vision transformer neural network structure consists of an image block embedding layer, 3 encoding layers and a convolutional layer.

[0037] The output of the image block embedding layer is connected to the input of the encoding layer, and the output of the encoding layer is connected to the input of the convolutional layer; the three encoding layers are connected in series;

[0038] The image block embedding layer is used to cut the image into patches of a fixed size of 10x10, and the dimension of the image embedding layer is set to 1024;

[0039] Encoding layer, including multi-head attention mechanism, the number of heads in each multi-head attention mechanism is 64, the dimension of each head in the multi-head attention mechanism is set to 64, and the dimension of the middle hidden layer of the multi-layer perceptron is 1024;

[0040] The convolution layer uses a 1x1 convolution kernel to change the number of channels of the feature map without changing the spatial dimension.

[0041] It should be noted that, in this article, relational terms such as first and second, etc. 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 terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0042] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A typhoon storm surge floodplain intelligent forecasting method based on Vision Transformer neural network, characterized in that: The steps of its intelligent forecasting method include: Obtaining continuous storm surge water level data and wind field data at the forecast time; the wind field data includes the component u of the wind speed in the east-west direction and the component v of the wind speed in the north-south direction; Preprocess the continuous storm surge water level data and wind field data at the forecast time, normalize the preprocessed data to the maximum value, and then divide them into training set and test set; Based on the training set, the normalized values ​​are input into the Vision transformer neural network for training to obtain the typhoon storm surge floodplain intelligent forecasting model; the test set is input into the typhoon storm surge floodplain intelligent forecasting model to adjust and optimize the typhoon storm surge floodplain to obtain the final typhoon storm surge floodplain intelligent forecasting model; The typhoon storm surge floodplain intelligent forecasting model is used to forecast the typhoon storm surge floodplain based on continuous storm surge water level data and wind field data at the forecast time.

2. According to claim 1, a typhoon storm surge floodplain intelligent forecasting method based on Vision Transformer neural network is characterized in that: The preprocessing is to combine the array of continuous storm surge water level data and the array of wind field data at the forecast time into one sample. The preprocessed data is four-dimensional data, and the four-dimensional data is data volume*number of channels*280*190.

3. The method for intelligent forecasting of typhoon storm surge floodplain based on Vision Transformer neural network according to claim 1, characterized in that: The maximum value normalization is to scale the data to between 0 and 1.

4. The method for intelligent forecasting of typhoon storm surge floodplain based on Vision Transformer neural network according to claim 1, characterized in that: The Vision transformer neural network structure consists of an image block embedding layer, 3 encoding layers and a convolutional layer; The output end of the image block embedding layer is connected to the input end of the encoding layer, and the output end of the encoding layer is connected to the input end of the convolutional layer; the three encoding layers are connected in series.

5. The method for intelligent forecasting of typhoon storm surge floodplains based on Vision Transformer neural network according to claim 4 is characterized in that: The image block embedding layer is used to cut the image into patches of a fixed size of 10x10, and the dimension of the image embedding layer is set to 1024; The encoding layer includes a multi-head attention mechanism, the number of heads in each multi-head attention mechanism is 64, the dimension of each head in the multi-head attention mechanism is set to 64, and the dimension of the middle hidden layer of the multi-layer perceptron is 1024; The convolution layer uses a 1x1 convolution kernel to change the number of channels of the feature map without changing the spatial dimension.

6. The method for intelligent forecasting of typhoon storm surge floodplain based on Vision Transformer neural network according to claim 1, characterized in that: The prediction steps of the typhoon storm surge floodplain intelligent forecasting model include: The image data in the training set is input into the image block embedding layer, which cuts the image data into patches of a fixed size of 10x10, and then uses a trainable linear projection matrix E to map the vectorized image blocks to a potential embedding space vector with D dimensions to obtain the embedding vector of the image blocks. Then, a corresponding D-dimensional position embedding vector generated for each image block is added to the embedding vector of the corresponding image block to obtain the final image patch embedding vector to preserve the spatial position information of the image block in the original image. The image patch embedding vector is input into the encoding layer, and the data is processed in parallel by a multi-head attention mechanism with a dimension of 64, focusing on learning different information dimensions, and then the nonlinear representation ability is enhanced by a multi-layer perceptron with a hidden layer dimension of 1024; After passing through three identical encoding layers, the data is input into the convolution layer and operated with a 1x1 convolution kernel to output the typhoon storm surge floodplain forecast.

Images