Storm surge space-time prediction method based on deep learning neural network

Abstract

The invention discloses a storm surge space-time prediction method based on a deep learning neural network, and relates to the technical field of data processing, and the method comprises the steps: constructing a space-time variable data set according to typhoon features and typhoon space-time features; the typhoon characteristics comprise typhoon paths, typhoon wind speeds, typhoon air pressures and storm surges corresponding to the typhoon paths; determining storm surge, typhoon wind speed and typhoon pressure of each time step as input parameters according to the time-space variable data set; inputting the input parameters into a pre-constructed deep learning neural network for training, so that the deep learning neural network predicts the storm surge of the next time step of each time step as an output parameter; and inputting a to-be-predicted input parameter into the trained deep learning neural network to obtain a predicted storm surge. According to the method, the typhoon features and the spatial-temporal features of the typhoon are fully integrated, the trained deep learning neural network is utilized to obtain the accurate predicted storm surge, and the capability and accuracy of predicting the spatial-temporal distribution of the storm surge are improved.

Description

Technical Field

[0001] This application relates to the technical field of data processing, and particularly to a spatio-temporal prediction method for storm surges based on deep learning neural networks. Background Art

[0002] Storm surge is an abnormal sea level rise phenomenon caused by atmospheric disturbances (such as strong winds and barometric changes), which seriously threatens the lives and property of coastal areas. With the intensification of global climate change, the occurrence frequency and intensity of storm surges are increasing day by day, making the disaster impacts caused by them more and more serious. To carry out storm surge disaster prevention and mitigation work efficiently and comprehensively, it is particularly important to achieve fast and accurate spatio-temporal forecasting of storm surges. Although traditional numerical models have achieved certain results in spatio-temporal forecasting of storm surges, due to their large consumption of computing resources and relatively slow computing speed, they are difficult to meet the requirements of real-time warning in terms of emergency response and rapid decision-making. In recent years, deep learning models have shown significant advantages in the forecasting work in the field of earth sciences, especially in terms of computing resource consumption and computing speed, providing a new solution idea for spatio-temporal forecasting of storm surges.

[0003] In the field of storm surge prediction, early deep learning-based methods mainly focused on time series prediction, mainly because the observational data was relatively scarce. Most of these studies relied on key meteorological parameters such as typhoon wind speed, wind direction, and barometric pressure as inputs, and used models such as artificial neural networks (ANN) and support vector machines (SVM) to predict storm surges at specific locations, and achieved satisfactory results to a certain extent. However, the existing solutions have the problem of lacking the ability to infer the spatial distribution and evolution pattern of target features. Summary of the Invention

[0004] In view of this, embodiments of this application provide a spatio-temporal prediction method for storm surges based on deep learning neural networks to improve the accuracy of predicting the spatial distribution and evolution pattern of storm surges.

[0005] One aspect of embodiments of this application provides a spatio-temporal prediction method for storm surges based on deep learning neural networks, and the method includes the following steps:

[0006] Construct a spatio-temporal variable data set according to typhoon characteristics and the spatio-temporal characteristics of typhoons; wherein, the typhoon characteristics include typhoon path, typhoon wind speed, typhoon barometric pressure, and the storm surge corresponding to the typhoon path;

[0007] Determine the storm surge, the typhoon wind speed, and the typhoon barometric pressure at each time step as input parameters according to the spatio-temporal variable data set;

[0008] Input the input parameters into a pre - constructed deep - learning neural network for training, so that the deep - learning neural network predicts the storm surge at the next time step for each time step as the output parameter;

[0009] Input the input parameters to be predicted into the trained deep - learning neural network to obtain the predicted storm surge.

[0010] In some embodiments, the step of calculating the typhoon air pressure includes the following steps:

[0011] Determine the air pressure at the calculation point in the typhoon area;

[0012] The expression for the air pressure at the calculation point is:

[0013] ;

[0014] Where, represents the air pressure at the calculation point; represents the peripheral air pressure of the typhoon; represents the central air pressure of the typhoon; is a parameter characterizing the typhoon system, calculated by an empirical formula and adjusted and corrected according to the radius of maximum wind speed; is the distance from the calculation point to the typhoon center.

[0015] In some embodiments, the step of calculating the typhoon wind speed includes the following steps:

[0016] Determine the wind speed distribution of the typhoon according to the air pressure distribution of the typhoon and the gradient wind relationship;

[0017] The expression for the wind speed distribution is:

[0018] ;

[0019] Where, is the wind speed distribution, is the Coriolis parameter, ; represents the angular velocity of the earth's rotation, represents the latitude, represents the air density; is the distance from the calculation point in the typhoon area to the typhoon center;

[0020] Determine the wind speed generated by the typhoon movement as the moving wind speed;

[0021] The expression for the moving wind speed is:

[0022] ;

[0023] Among them, is the moving wind speed, and respectively represent the eastward component and the northward component of the moving speed of the typhoon center;

[0024] Determine the wind speed of the symmetric typhoon model according to the wind speed distribution and the moving wind speed;

[0025] The expression of the wind speed of the symmetric typhoon model is:

[0026] ;

[0027] Among them, is the wind speed of the symmetric typhoon model, is the angle between the line connecting the calculation point and the typhoon center and the due east direction, is the angle between the gradient wind and the surface wind, and are correction factors;

[0028] Determine the typhoon wind speed at each calculation point in the typhoon area according to the wind speed of the symmetric typhoon model;

[0029] The expression of the typhoon wind speed is:

[0030] ;

[0031] Among them, is the typhoon wind speed; is an intermediate parameter, ; represents the background wind speed of the 10-meter wind speed component from the ERA5 reanalysis dataset; and respectively represent the threshold distances selected and determined through verification experience.

[0032] In some embodiments, the steps of determining the storm surge corresponding to the typhoon path include the following steps:

[0033] Input the typhoon wind speed and the typhoon air pressure into the FVCOM model to obtain the storm surge in the format of triangular grid data;

[0034] Use the inverse distance weighted interpolation method to convert the storm surge in the format of triangular grid data into the storm surge in the format of rectangular grid data as the storm surge corresponding to the typhoon path.

[0035] In some embodiments, before inputting the input parameters into a pre-constructed deep learning neural network for training, the method further includes the following steps:

[0036] Construct the deep learning neural network based on the vision transformer and incorporate the adaptive Fourier neural operator;

[0037] Based on the architecture of the vision transformer, use the adaptive Fourier neural operator to replace and optimize the self-attention algorithm in the deep learning neural network, and perform spatial mixing operations on spatial features and channel features.

[0038] In some embodiments, inputting the input parameters to be predicted into the trained deep learning neural network to obtain the predicted storm surge includes the following steps:

[0039] Take the storm surge, typhoon wind speed, typhoon air pressure, and the corresponding time step in the input parameters corresponding to each time step as one channel respectively, and then divide the input parameters into several small data blocks;

[0040] Use the encoder in the deep learning neural network to embed each small data block into a higher-dimensional space, and then apply position encoding to generate tokens corresponding to each small data block;

[0041] Use the adaptive Fourier neural operator in the trained deep learning neural network to mix each token, and then determine the predicted storm surge according to the mixed tokens.

[0042] In some embodiments, using the adaptive Fourier neural operator in the trained deep learning neural network to mix each token, and then determining the predicted storm surge according to the mixed tokens includes the following steps:

[0043] Use the adaptive Fourier neural operator to perform a fast Fourier transform to transform the small data blocks corresponding to each token into the frequency domain;

[0044] Use the adaptive Fourier neural operator to set weights for the corresponding tokens through the diagonal weight matrix and soft threshold operation of the small data blocks in the frequency domain;

[0045] Perform a non-linear mapping on each small data block according to the weights to improve the sparsity of each small data block;

[0046] Repeat the token processing steps a set number of times to obtain the final tokens. The token processing steps include: for the spatial mixing of each token, use a multi-layer MLP for transformation to obtain the mixed tokens; convert the mixed tokens back to the spatial domain through an inverse Fourier transform, and then perform a residual connection; for the channel mixing of each token, use a single-layer MLP for transformation to achieve non-linear mapping of features;

[0047] Using a decoder to embed the final token to the reconstruction time to obtain the predicted storm surge.

[0048] Another aspect of the embodiments of the present application also provides a storm surge spatio-temporal prediction device based on a deep learning neural network. The device includes:

[0049] A dataset construction unit for constructing a spatio-temporal variable dataset according to typhoon characteristics and spatio-temporal characteristics of typhoons; wherein, the typhoon characteristics include typhoon path, typhoon wind speed, typhoon pressure, and the storm surge corresponding to the typhoon path.

[0050] An input parameter determination unit for determining the storm surge, the typhoon wind speed, and the typhoon pressure at each time step as input parameters according to the spatio-temporal variable dataset.

[0051] A neural network training unit for inputting the input parameters into a pre-constructed deep learning neural network for training, so that the deep learning neural network predicts the storm surge at the next time step of each time step as an output parameter.

[0052] A storm surge prediction unit for inputting the input parameters to be predicted into the trained deep learning neural network to obtain a predicted storm surge.

[0053] Another aspect of the embodiments of the present application also provides an electronic device, including a processor and a memory;

[0054] The memory is used to store a program;

[0055] The processor executes the program to implement the method described in any one of the above.

[0056] Another aspect of the embodiments of the present application also provides a computer-readable storage medium. The storage medium stores a program, and the program is executed by a processor to implement the method described in any one of the above.

[0057] The present application at least includes the following beneficial effects:

[0058] This application can construct a spatio-temporal variable dataset based on typhoon characteristics and the spatio-temporal characteristics of typhoons; among them, typhoon characteristics include typhoon track, typhoon wind speed, typhoon pressure, and storm surge corresponding to the typhoon track; determine the storm surge, typhoon wind speed, and typhoon pressure at each time step as input parameters according to the spatio-temporal variable dataset; input the input parameters into a pre-constructed deep learning neural network for training, so that the deep learning neural network predicts the storm surge at the next time step of each time step as the output parameter; input the input parameters to be predicted into the trained deep learning neural network to obtain the predicted storm surge. This application fully integrates typhoon characteristics and the spatio-temporal characteristics of typhoons, and then uses the trained deep learning neural network to obtain accurate predicted storm surges, improving the ability and accuracy of predicting the spatio-temporal distribution of storm surges. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0060] Figure 1 It is a schematic flowchart of a semantic segmentation method based on mutual relationship knowledge distillation provided by an embodiment of the present application;

[0061] Figure 2 It is a schematic diagram of the wind speed distribution of the synthetic typhoon track provided by an embodiment of the present application;

[0062] Figure 3 It is a framework diagram of the DSSP model provided by an embodiment of the present application;

[0063] Figure 4 It is a schematic diagram of the track of typhoon "Phrawhin" provided by an embodiment of the present application;

[0064] Figure 5 It is a schematic diagram of the air pressure and wind field changes of typhoon "Phrawhin" provided by an embodiment of the present application;

[0065] Figure 6 It is a comparison diagram of the storm surge prediction results of the DSSP model and the FVCOM model provided by an embodiment of the present application;

[0066] Figure 7 It is a schematic diagram of the spatial average RMSE and CORR between the results of the DSSP model and the target results (FVCOM model results) provided by an embodiment of the present application;

[0067] Figure 8A comparison chart of storm surge data observed in the Yangtze Estuary and coastal areas and the prediction results of the FVCOM and DSSP models provided by the embodiments of this application;

[0068] Figure 9 A schematic diagram of the geographical distribution of each site provided by the embodiments of this application;

[0069] Figure 10 A comparison chart of the spatial average RMSE and CORR between the DSSP prediction results and the target (FVCOM results) provided by the embodiments of this application;

[0070] Figure 11 A structural block diagram of a storm surge spatio-temporal prediction device based on a deep learning neural network provided by the embodiments of this application. Detailed implementation manners

[0071] In order to make the objectives, technical solutions and advantages of this application clearer, the following further elaborates on this application in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not used to limit this application.

[0072] Before elaborating on the embodiments of this application in detail, first explain the terms and some related technologies that may be involved in the embodiments of this application as follows:

[0073] ANN (Artificial Neural Networks): Artificial Neural Networks, a machine learning model inspired by biological neural networks, consisting of multiple artificial neurons, capable of learning complex mapping relationships through training.

[0074] MLP (Multi-Layer Perceptron): Multi-Layer Perceptron, a feedforward neural network composed of an input layer, a hidden layer, and an output layer. It uses fully connected layers to transmit information and introduces non-linearity through activation functions. MLP is optimized through backpropagation and is widely used in classification, regression, and feature extraction tasks. It is a typical structure of ANN.

[0075] SVM (Support Vector Machines): Support Vector Machines, a supervised learning algorithm mainly used for classification and regression analysis. It constructs a hyperplane to maximize the interval between classes, has good generalization ability, and is suitable for high-dimensional data and small-sample problems.

[0076] CONV-LSTM (Convolutional Long Short-Term Memory): A model that combines a Convolutional Neural Network (CNN) and a Long Short-Term Memory network (LSTM), capable of capturing both spatial and temporal features simultaneously, and is widely used in spatio-temporal sequence prediction.

[0077] FVCOM (Finite Volume Coastal Ocean Model): A three-dimensional ocean numerical model that uses the finite volume method to solve the shallow water equations on an unstructured triangular grid. Its core equations include the momentum equation, the mass continuity equation, and the equations for temperature, salinity, and density. Due to the use of triangular grids, FVCOM can perform local refinement in coastal areas, providing a smoother coastline transition, and is particularly suitable for simulating complex coastlines. This model is widely used in storm surge simulations.

[0078] AFNO (Adaptive Fourier Neural Operator): The Adaptive Fourier Neural Operator is an efficient token mixer. Based on the Fourier Neural Operator (FNO), AFNO performs global convolution in the Fourier domain, bypassing the computational complexity bottleneck of self-attention, and achieving quasi-linear computation and linear memory consumption. To improve efficiency, AFNO uses a block diagonal structure to constrain channel mixing, adaptive weight sharing to enhance generalization, and selects key frequency patterns through frequency domain sparsification to reduce computational costs while retaining the ability to model global information.

[0079] ViT (Vision Transformer): A Vision Transformer, a deep learning model based on the Transformer architecture, which divides an image into small patches and uses the self-attention mechanism to capture global features. Compared with CNN, it performs excellently when trained on large-scale data and is widely used in tasks such as image classification and object detection.

[0080] DSSP (Deep Learning Storm Surge Prediction): A deep learning model for storm surge spatio-temporal prediction that uses ViT as the backbone and AFNO.

[0081] T-tide: A tidal analysis tool based on harmonic analysis, which can be used to extract the main tidal components from tide level or current data for tidal prediction and tidal characteristic analysis.

[0082] RMSE (Root Mean Square Error): The correlation coefficient reflects the linear relationship between the predicted value and the actual value, and its calculation formula is as follows:

[0083] ;

[0084] Among them, and respectively represent the target value and the predicted value, and are the means of the target value and the predicted value respectively.

[0085] CORR (Correlation Coefficient): The calculation formula of the root mean square error is as follows:

[0086] ;

[0087] Among them, is the total number of predicted samples.

[0088] The regional changes in storm surges provide valuable information for many industries (such as the shipping industry). Therefore, this application further explores a deep learning model capable of predicting the spatio-temporal changes in storm surges. Currently, there are also methods for spatio-temporal prediction of storm surges through models such as CONV-LSTM, but these methods mainly rely on historical meteorological and tide level data for prediction and fail to fully integrate real-time typhoon information. In addition, the relatively scarce measured historical typhoon path data makes data-driven deep learning methods face the problem of insufficient data. These aspects limit the training effect and prediction accuracy of the model.

[0089] Referring to Figure 1 , the embodiments of this application provide a spatio-temporal prediction method for storm surges based on a deep learning neural network, specifically including the following steps S100 to S130:

[0090] S100: Construct a spatio-temporal variable data set according to the typhoon characteristics and the spatio-temporal characteristics of the typhoon; among them, the typhoon characteristics include the typhoon path, typhoon wind speed, typhoon pressure, and the storm surge corresponding to the typhoon path;

[0091] S110: Determine the storm surge, the typhoon wind speed, and the typhoon pressure at each time step as input parameters according to the spatio-temporal variable data set;

[0092] S120: Input the input parameters into a pre-constructed deep learning neural network for training, so that the deep learning neural network predicts the storm surge at the next time step of each time step as the output parameter;

[0093] S130: Input the input parameter to be predicted into the trained deep learning neural network to obtain a predicted storm surge.

[0094] Optionally, the step of calculating the typhoon air pressure includes the following steps:

[0095] Determine the air pressure at the calculation point in the typhoon area;

[0096] The expression for the air pressure at the calculation point is:

[0097] ;

[0098] Where, represents the air pressure at the calculation point; represents the peripheral air pressure of the typhoon; represents the central air pressure of the typhoon; is a parameter characterizing the typhoon system, is calculated through an empirical formula and adjusted and corrected according to the radius of maximum wind speed; is the distance from the calculation point to the typhoon center.

[0099] Optionally, the step of calculating the typhoon wind speed includes the following steps:

[0100] Determine the wind speed distribution of the typhoon according to the air pressure distribution of the typhoon and the gradient wind relationship;

[0101] The expression for the wind speed distribution is:

[0102] ;

[0103] Where, is the wind speed distribution, is the Coriolis parameter, ; represents the angular velocity of the Earth's rotation, represents the latitude, represents the air density; is the distance from the calculation point to the typhoon center in the typhoon area;

[0104] Determine the wind speed generated by the typhoon movement as the moving wind speed;

[0105] The expression for the moving wind speed is:

[0106] ;

[0107] Where, is the moving wind speed, and respectively represent the eastward component and northward component of the typhoon center movement speed;

[0108] Determine the wind speed of the symmetric typhoon model according to the wind speed distribution and the moving wind speed;

[0109] The expression for the wind speed of the symmetric typhoon model is:

[0110] ;

[0111] where, is the wind speed of the symmetric typhoon model, is the angle between the line connecting the calculation point and the typhoon center and the due east direction, is the angle between the gradient wind and the surface wind, and are correction factors;

[0112] Determine the typhoon wind speed at each calculation point in the typhoon area according to the wind speed of the symmetric typhoon model;

[0113] The expression for the typhoon wind speed is:

[0114] ;

[0115] where, is the typhoon wind speed; is an intermediate parameter, ; represents the background wind speed of the 10-meter wind speed component from the ERA5 reanalysis dataset; and respectively represent the threshold distances selected and determined through verification experience.

[0116] Optionally, the steps for determining the storm surge corresponding to the typhoon path include the following steps:

[0117] Input the typhoon wind speed and the typhoon pressure into the FVCOM model to obtain the storm surge in the triangular grid data format;

[0118] Convert the storm surge in the triangular grid data format into the storm surge in the rectangular grid data format by using the inverse distance weighted interpolation method as the storm surge corresponding to the typhoon path.

[0119] Optionally, before inputting the input parameters into the pre-constructed deep learning neural network for training, the method further includes the following steps:

[0120] Construct the deep learning neural network based on the vision transformer and incorporating the adaptive Fourier neural operator;

[0121] Based on the architecture of the vision transformer, the adaptive Fourier neural operator is used to replace and optimize the self-attention algorithm in the deep learning neural network to perform spatial mixing operations on spatial features and channel features.

[0122] Optionally, inputting the input parameter to be predicted into the trained deep learning neural network to obtain a predicted storm surge includes the following steps:

[0123] Taking the storm surge, typhoon wind speed, typhoon air pressure, and corresponding time step of the corresponding time step in the input parameter as one channel respectively, and then dividing the input parameter into several small data blocks;

[0124] Using the encoder in the deep learning neural network to embed each small data block into a higher-dimensional space, and then applying position encoding to generate tokens corresponding to each small data block;

[0125] Using the adaptive Fourier neural operator in the trained deep learning neural network to perform mixing processing on each token, and then determining the predicted storm surge according to the mixed tokens.

[0126] Optionally, using the adaptive Fourier neural operator in the trained deep learning neural network to perform mixing processing on each token, and then determining the predicted storm surge according to the mixed tokens includes the following steps:

[0127] Using the adaptive Fourier neural operator to perform a fast Fourier transform to transform the small data block corresponding to each token into the frequency domain;

[0128] Using the adaptive Fourier neural operator to set weights for the corresponding tokens through the diagonal weight matrix and soft threshold operation of the small data block in the frequency domain;

[0129] Performing a non-linear mapping on each small data block according to the weights to improve the sparsity of each small data block;

[0130] Repeating the token processing step a set number of times to obtain the final token. The token processing step includes: performing spatial mixing on each token, using a multi-layer MLP for transformation to obtain the mixed token; converting the mixed token back to the spatial domain through an inverse Fourier transform, and then performing a residual connection; performing channel mixing on each token, using a single-layer MLP for transformation to achieve non-linear mapping of features;

[0131] Using the decoder to embed the final token to the reconstruction time to obtain the predicted storm surge.

[0132] Next, specific application examples will be combined to introduce and explain the solutions of the embodiments of the present application in detail.

[0133] Specifically, this embodiment may include the following technical solutions:

[0134] 1. Dataset construction.

[0135] (1) Calculation area:

[0136] Optionally, the detailed geographical coordinates of the calculation area are from 117.40 °E to 137.19 °E and from 19.90 °N to 42.00 °N. This area extends eastward from Zhenjiang to the estuary and further covers the offshore area, with a water depth range of 40 to 50 meters. This area is located in the subtropical monsoon climate zone, where typhoons occur frequently in summer and autumn, often triggering severe storm surges.

[0137] (2) Typhoon paths:

[0138] The typhoon path data used for model training includes measured paths and synthetic paths. The measured typhoon paths are sourced from the China Meteorological Administration's Tropical Cyclone Best Track Dataset, which records detailed information on tropical cyclones in the Northwest Pacific and South China Sea regions since 1949. This dataset covers the trajectory and intensity data of tropical cyclones and is updated every six hours. Between 1979 and 2024, 155 typhoons that existed in the calculation area for more than 24 hours were selected for analysis. The intensities of the above 155 tropical cyclones are divided into six grades based on the maximum near-surface wind speed, and this classification standard refers to the national standard for tropical cyclone classification issued by the China Meteorological Administration.

[0139] In addition, the synthetic typhoon paths, as an important supplement, provide a wider range of paths and intensities. These paths simulate typhoon processes under different atmospheric conditions and climate backgrounds, effectively making up for the limitations of observational data in terms of coverage and sample size. The synthetic typhoon paths are generated by a statistical-deterministic tropical cyclone model based on reanalysis data and global climate model outputs. These paths include 2,466 typhoon paths that existed in the calculation area for more than 24 hours and affected the study area. To ensure the representativeness of the dataset and its consistency with the actual typhoon intensity distribution, the wind speed distributions of these paths have been adjusted according to historical typhoon records. Specifically, the occurrence frequencies of each intensity grade have been calibrated to reflect the observed statistical data characteristics, emphasizing the high incidence of tropical depressions and tropical storms, while also considering the relatively rare but highly impactful phenomenon of super typhoons. Exemplarily, Figure 2 is a schematic diagram of the wind speed distribution of the synthetic typhoon path.

[0140] (3) Symmetric typhoon model:

[0141] The calculation of the wind field and the pressure field adopts a symmetric typhoon model, which is a simplified representation method of the typhoon wind field structure and is used to provide meteorological data for numerical models and deep learning models. This model assumes that the typhoon has symmetric wind field characteristics and can accurately reflect the wind speed distribution in different directions. The main parameters of the model include the position of the typhoon, the central pressure ( ), the maximum wind speed near the typhoon center ( ), and the radius of the maximum wind speed ( ), etc.

[0142] The formula for calculating the pressure at the calculation point ( ) is as follows:

[0143] (1)

[0144] Where, represents the peripheral pressure of the typhoon (set as the standard atmospheric pressure), is a parameter characterizing the typhoon system (calculated by an empirical formula and corrected according to adjustment), is the distance from the calculation point to the typhoon center.

[0145] The wind speed distribution of the cyclone ( ) is derived based on the gradient wind relationship and referring to the pressure distribution:

[0146] (2)

[0147] (3)

[0148] Where, is the Coriolis parameter, represents the angular velocity of the Earth's rotation, represents the latitude, represents the air density. The wind speed generated by the typhoon movement is calculated as follows:

[0149] (4)

[0150] Where, and respectively represent the eastward component and the northward component of the typhoon center's movement speed. The wind speed represented by the symmetric model can be expressed as:

[0151] (5)

[0152] Where, is the angle between the line connecting the calculation point and the typhoon center and the due east direction, is the angle between the gradient wind and the surface wind, and is the correction coefficient, both taking 0.8. The typhoon wind speed ( ) can be calculated by the following formula:

[0153] (6)

[0154] (7)

[0155] where represents the background wind speed of the 10-meter wind speed component from the ERA5 reanalysis dataset; and respectively represent the threshold distances selected and determined through verified experience, taking values of 60 km and 200 km respectively.

[0156] (4) Numerical model:

[0157] The numerical model is used to calculate the storm surge corresponding to a given typhoon path. In this study, the The Finite-Volume Coastal Ocean Model (FVCOM) is adopted, which is a three-dimensional ocean numerical model based on an unstructured triangular grid with a free surface. The core equations of this model include the momentum equation, the mass continuity equation, and the equations for temperature, salinity, and density. The FVCOM model uses triangular grids, which can be locally refined in coastal areas to provide a smoother coastline transition, so it is particularly suitable for simulating complex coastlines. FVCOM has been widely applied in storm surge modeling and has performed well in simulating large-scale storm surges and local impacts in different coastal areas.

[0158] The grid structure of the FVCOM model adopted in this embodiment is carefully designed. Specifically, the grid resolution in the offshore area is set to 0.5 to 5 km to balance the calculation efficiency and accuracy, and it extends to the Datong tidal level measurement station at the river boundary. The FVCOM model has been continuously developed and widely verified. For each typhoon event, the starting time of the numerical simulation is set to seven days before the typhoon occurs to eliminate the initial numerical perturbation. To focus on the mapping relationship between meteorological information and storm surge, the influence of astronomical tides is excluded in the numerical model.

[0159] (5) Dataset composition:

[0160] For each typhoon case, the symmetric typhoon model provides the eastward and northward wind speed components and air pressure, which are used as meteorological input variables for the deep learning model. These variables are also the inputs of the FVCOM model, which calculates and outputs storm surge data for the deep learning model to form a dataset containing four spatio-temporal variables. This dataset is stored in a four-dimensional structural format with dimensions of time, feature (variable), latitude, and longitude. It should be noted that the time resolution for each typhoon event is 1 hour, and the total data length is approximately 3 to 5 days.

[0161] The FVCOM model uses an unstructured grid with a flexible spatial resolution. However, the deep learning model requires rectangular grid data. Therefore, the inverse distance weighted interpolation method is used to convert the triangular grid data of FVCOM into a 440×400 rectangular grid with a resolution of 0.05°. This interpolation method ensures a smooth transition between land and sea and reduces potential boundary effects.

[0162] In addition to the dataset used to train the deep learning model, the observed time series data of multiple tide gauge stations are collected to evaluate the performance of the deep learning model in storm surge forecasting. These tide gauge stations cover the main coastal and estuarine areas within the study area and can comprehensively verify the deep learning model. The original tide data contains the combined effects of astronomical tides and storm surges. Therefore, in this embodiment, T-tide is used to perform harmonic analysis on the original tide data. Through the results of harmonic analysis, this embodiment subtracts the calculated astronomical tide component from the original tide observations to isolate the storm surge.

[0163] 2. Model establishment.

[0164] The DSSP (Deep Learning Storm Surge Prediction) model constructed in this embodiment, namely the deep learning neural network, utilizes an architecture that combines the Adaptive Fourier Neural Operator (AFNO) and the Vision Transformer (ViT). ViT can effectively capture the spatial and channel features in the data. Although ViT exhibits strong feature capture capabilities, the spatial mixing operation implemented through the self-attention mechanism has a quadratic computational complexity in terms of the number of input tokens. For high-resolution input parameters, such as large-scale high-resolution ocean data, this quadratic complexity significantly increases the computational requirements. To address the computational challenges in high-resolution scenarios, the DSSP model adopts the AFNO technique to replace the self-attention mechanism.

[0165] Figure 3Shows the framework of the DSSP model of this embodiment. The model takes the values of storm surge, wind speed component, and sea level pressure at time t as inputs and the storm surge at time t+1 as the output. The input parameters constitute four channels of the model input (i.e., c = 4). The input parameters are split into small data chunks, each of which has dimensions. Subsequently, these small data chunks are embedded into a higher-dimensional space, and positional encoding is applied to generate tokens (one-dimensional matrix vectors). Next, AFNO is used to mix the tokens.

[0166] Specifically, AFNO first performs a fast Fourier transform to transform the small data chunks corresponding to each token into the frequency domain. In the frequency domain, AFNO weights the tokens through the diagonal weight matrix of the small data chunks and a soft threshold operation, and promotes sparsity.

[0167] Two layers of multi-layer perceptrons (MLPs) are applied for non-linear mapping:

[0168] (8)

[0169] where, represents the token, is the ReLU activation function, , and are the shared weights of each block, represents the total number of blocks in the block diagonal matrix, represents the calculation of the th block matrix. Subsequently, a soft threshold operation is performed to promote sparsity:

[0170] (9)

[0171] where, is the adjustment parameter for controlling sparsity. Subsequently, a multi-layer MLP is used to transform the spatial mixing of each token. The mixed tokens are then transformed back to the spatial domain through an inverse Fourier transform (IFFT), followed by a residual connection. Subsequently, a single-layer MLP is used to transform the channel mixing of each token to achieve non-linear mapping of features. The above steps of spatial mixing and channel mixing of each token are repeated times. Finally, the decoder reconstructs the storm surge at time based on the final token embedding.

[0172] The Fourier neural operator (FNO) can simplify the global convolution in the spatial domain to a multiplication operation in the Fourier domain:

[0173] (10)

[0174] wherein, represents a token; , is the position of the index block; is the weight matrix. Specifically, AFNO introduces a block diagonal structure in FNO, and divides the original weight matrix ( is the dimension of the token) into weight blocks of size . This block diagonal structure reduces the mixing complexity from to , where is the number of blocks or tokens. Therefore, AFNO achieves quasi-linear computational complexity and significantly reduces memory usage.

[0175] 3. Model training.

[0176] A dataset containing 155 observed typhoon tracks and 2,466 synthetic typhoon tracks is divided into a training set, a validation set, and a test set in a ratio of 7:2:1. Before the division, the typhoon tracks are randomly shuffled to ensure that each subset contains typhoons of different intensities and includes a balanced distribution of observed and synthetic typhoon tracks.

[0177] The training process is carried out on a single Nvidia A800 GPU, with a total running time of about 48 hours to complete 80 epochs of training. This hardware configuration and optimization strategy ensure efficient training while maintaining stable performance. The detailed hyperparameter settings are summarized in Table 1. The parameters in Table 1 are fine-tuned through multiple sensitivity tests to achieve the best balance between computational efficiency and prediction accuracy.

[0178] Table 1

[0179] Hyperparameter Setting Block size 2 Embedding dimension 512 Model depth 6 Number of AFNO blocks 4 Optimizer Adam Learning rate <![CDATA[5×10 -3 > Batch size 4 Number of training epochs 80

[0180] The model uses the Adam optimizer and combines a learning rate decay strategy, gradually reducing the learning rate to a preset minimum value using a cosine decay schedule. This method improves the stability of training and alleviates the overfitting problem. The embedding dimension determines the length of the vector into which each block is transformed before being input into AFNO. Although a higher embedding dimension can capture more features, it also increases the computational cost. Therefore, the embedding dimension is set to 512.

[0181] In AFNO, the weight matrix is structured in a block - diagonal form and divided into sub - blocks for independent calculation. The model is configured with 4 blocks to improve accuracy and parallel efficiency. The depth of the model refers to the number of mixing layers. Increasing the number of layers allows the model to learn more complex features, but may also lead to the vanishing gradient problem. The depth of the mixing layer is set to 6 layers to optimize feature learning while maintaining gradient stability.

[0182] Next, more specific embodiments will be described.

[0183] Using the deep - learning model (deep - learning neural network) of this embodiment for the operational forecasting and early warning of the storm surge disaster of Typhoon "Praison" that landed on September 19, 2024, the forecasting results have high reliability. The following are the implementation schemes:

[0184] 1. Spatiotemporal prediction:

[0185] "Praison" is the 14th typhoon in the 2024 Pacific typhoon season, with a maximum central wind speed of 32.1 m / s, belonging to the tropical storm intensity. As a medium - strength typhoon, "Praison" has typical conditions to trigger storm surges. Therefore, it is used as a representative case to evaluate the prediction performance of the DSSP model and demonstrate its wide applicability. "Praison" formed in the northwestern Pacific on the evening of September 15, 2024 (UTC + 8) and developed into a severe tropical storm on the morning of September 19. At 21:45 on September 19, "Praison" made a coastal landfall. The schematic diagram of the path of Typhoon "Praison" is as Figure 4 shown.

[0186] The initial prediction time of the model is set at 09:00 on September 18, 2024 (UTC + 8). At the start of the prediction, observed typhoon information was collected, and the initial storm - surge field was calculated through the FVCOM model. During the subsequent prediction process, the future typhoon path was provided by the prediction of meteorological agencies. Subsequently, the symmetric typhoon model was used to synchronously calculate the predicted pressure and wind fields as the input of the DSSP model. It should be noted that the input storm - surge information comes from the prediction of the previous step of the DSSP model and does not rely on numerical models, thus enabling fast step - by - step storm - surge prediction. This prediction covers the spatiotemporal evolution of the next 48 hours with a time step of 1 hour.

[0187] Figure 5 Shows the changes in the wind field and pressure field 12 hours before the landfall of Typhoon "Praison". As "Praison" approaches the landing point, obvious dynamic changes can be observed. The typhoon path and the sea - level pressure field clearly show the development of "Praison". 24 hours after the initial prediction time, the influence range of the typhoon is the largest, and the cyclone center is located southeast of the Yangtze River Estuary ( Figure 5a). After 30 hours, the typhoon approached the coastal area and maintained its intensity as a severe tropical cyclone ( Figure 5 b). After 36 hours, the typhoon made landfall in a certain coastal area ( Figure 5 c). At the time of landfall, the maximum wind speed near the typhoon center was 23 m / s, and the minimum central pressure was 99.5 kPa.

[0188] The comparison of the forecast results between the DSSP model and the FVCOM model shows that the DSSP model can effectively capture the development of the storm surge caused by the typhoon and better predict the area and time of the storm surge occurrence ( Figure 6 ). Specifically, 24 hours after the initial time, both models predicted a water level rise in the northwest direction of the typhoon center ( Figure 6 a, d). At the same time, both models also predicted a water level drop in the southwest direction of the typhoon center. The maximum storm surge predicted by the FVCOM model was 0.10 m, while the maximum storm surge predicted by the DSSP model was 0.24 m. The storm surge amplitude predicted by the DSSP model was slightly higher than that of the FVCOM model. After 30 hours, both models predicted that the storm surge was advancing significantly towards the coastal area ( Figure 6 b, e). The difference in the maximum storm surge between the two models was relatively small. The maximum storm surge predicted by the FVCOM model was 0.32 m, while the maximum storm surge predicted by the DSSP model was 0.26 m. The trends of the two models in the water level drop area were consistent. Although the spatial distribution of the drop area was different, the difference in the water level drop amplitude between the two was small, with a maximum difference of only 0.03 m. During the landing stage, the difference in the maximum storm surge between the two was 0.34 m. The storm surge intensity of the DSSP model was slightly lower than that of the FVCOM model, but the storm surge areas predicted by the two were similar in scope and shape ( Figure 6 c, f).

[0189] The above results indicate that although there are some differences in the storm surge intensity prediction between the two models, the overall distribution and trend of the storm surge are similar. It is worth noting that the DSSP model better controlled the cumulative error within the first 40 hours, which is of great significance for formulating countermeasures. Compared with the FVCOM model, the spatial average RMSE of the DSSP model remained below 0.2 m, and the CORR exceeded 0.9 ( Figure 7 ). Although the cumulative error gradually increased over time, the overall prediction performance was still reliable, with the RMSE remaining within 0.3 m and the CORR continuously remaining above 0.7.

[0190] 2. Time series verification:

[0191] To further verify the prediction results, in this embodiment, the measured data of multiple tide gauge stations are used for time series verification. These stations are distributed along the coastal areas. The DSSP model successfully captured the main peaks of the storm surge, including the timing and amplitude. Specifically, the DSSP model predicted that the storm surge at some stations exceeded 1.0 m ( Figure 8 for a, b, c). At these three stations, the RMSE between the DSSP model prediction and the FVCOM model output remained within 0.26 m. In addition, the time series verification also showed that the DSSP model was able to accurately predict the spatial differences in the storm surge intensity between different stations. For example, when a typhoon landed near the Yangtze River Estuary, the storm surge intensity at the Yangtze River Estuary station was significantly higher than that in some coastal areas. Additionally, compared with the FVCOM model, the DSSP model provided a more accurate prediction of the storm surge peak timing. The maximum storm surge value predicted by the DSSP model was closer to the measured value ( Figure 8 for a, b, c), which is of great significance for disaster response. Among them, Figure 8 shows the comparison of the storm surge data observed at the Yangtze River Estuary (a, b, c) and coastal areas (d, e, f) with the prediction results of the FVCOM and DSSP models. The RMSE and CORR are calculated based on the results of the DSSP model and the FVCOM model. Exemplarily, Figure 9 is a schematic diagram of the geographical distribution of each station.

[0192] To better demonstrate the overall performance of the DSSP model at these stations, this embodiment conducted a statistical analysis (Table 2). The first two columns of Table 2 show the 48-hour average error between the DSSP model and the FVCOM model, and the last two columns compare the errors of these two models relative to the observed maximum storm surge level. During the 48-hour forecast period, the average RMSE between the DSSP model and the FVCOM model at the observation stations was approximately 0.15 m, and the average correlation coefficient (CORR) was 0.94. In addition, in terms of the maximum storm surge level, the average error of the DSSP model was approximately 0.18 m, which was 45% lower than that of the FVCOM model. This result indicates that the DSSP model in this embodiment has a high prediction accuracy in capturing the storm surge.

[0193] Table 2

[0194]

[0195] In summary, this embodiment proposes a deep learning spatio-temporal prediction scheme for storm surges (DSSP). The DSSP model of this embodiment can effectively capture the complex spatio-temporal dynamics of storm surges, and while maintaining high prediction accuracy, it achieves a significant improvement in computational efficiency. The computational efficiency far exceeds that of traditional numerical models. It only takes 15 seconds to complete a 48-hour storm surge forecast, with a 99.58% improvement in computational efficiency. In terms of the prediction of storm surge peaks, the average error between the storm surge prediction of the DSSP model at the measuring stations and the measured data is smaller, with an average error of less than 0.18 meters, which is 45% lower than that of the FVCOM model.

[0196] In this embodiment, the introduction of synthetic typhoon tracks into the training dataset significantly improves the model performance. The experimental results show that the addition of synthetic typhoon data reduces the RMSE of the DSSP model of this embodiment by more than 50% compared with the FVCOM model, and at the same time, the correlation coefficient (CORR) increases from 0.87 to 0.94. In addition, the DSSP model combines real-time typhoon tracks and meteorological data to ensure that the prediction results can be dynamically updated and adapted to changing environmental conditions. By eliminating the influence of astronomical tides, the DSSP model can accurately extract storm surges and establish a more accurate relationship between meteorological variables and storm surge dynamics. This embodiment provides a reliable solution for the operational forecasting of storm surges, making significant progress in both prediction accuracy and computational efficiency, and providing important technical support for disaster prevention and mitigation work.

[0197] The beneficial effects of this embodiment include:

[0198] 1. Forecasting efficiency:

[0199] The prediction efficiency of the DSSP model in this embodiment far exceeds that of numerical models (see Table 3). For the same typhoon prediction task, the FVCOM model requires more than 60 minutes using 30 Intel(R) Xeon(R) Platinum 8358 CPU cores. While the DSSP model only takes 15 seconds and can complete the prediction using one Nvidia A800 GPU. Prediction efficiency is crucial for the public and relevant departments because it can significantly improve the emergency response speed to storm surge disasters, thus ultimately reducing the losses caused by storm surges.

[0200] Table 3

[0201] Model FVCOM DSSP (in this embodiment) Computing device type CPU GPU Computing resources Intel(R) Xeon(R) Platinum 8358 CPU Nvidia A800 GPU Computing time 60 minutes 15 seconds

[0202] 2. Improvement of model performance by expanding the dataset:

[0203] For data-driven deep learning models, the impact of the dataset is crucial. To test the effectiveness of the dataset in this embodiment, a comparative test using different datasets was conducted. In this experiment, the DSSP model was trained only using the dataset containing the measured typhoon tracks in this embodiment, while keeping the hyperparameters unchanged. Then, the trained model was applied to the prediction of Typhoon Phanfone. The test results show that for the model trained using only the dataset containing the measured typhoon tracks, during the period from 24 hours to 36 hours after the initial time, the spatially averaged RMSE exceeds 0.6 m and the CORR is lower than 0.6 ( Figure 10 ), while for the model trained using the dataset with synthetic tracks added, the RMSE and CORR are lower than 0.2 m and higher than 0.9, respectively. In addition to the comparison of these detailed values, the pattern of the metric curve changing over time clearly shows that in the case of only using the dataset of measured tracks, the control of the cumulative error is poor. This indicates that the synthetic typhoon paths help the model learn more diverse and comprehensive typhoon path characteristics, which is beneficial to the generalization ability of the DSSP model. Among them, Figure 10 shows the comparison of the spatially averaged RMSE and CORR between the DSSP prediction results and the target (FVCOM results). The solid line represents the results using the dataset in this embodiment, and the dashed line represents the results using only the measured typhoon paths.

[0204] 3. Incorporating real-time typhoon paths:

[0205] In this embodiment, the latest forecast typhoon tracks provided by the meteorological department are used to construct timely wind speed and pressure fields, which are then used as known input parameters for the current step. This method enables the model to incorporate the latest typhoon path information and not rely solely on historical data, thereby improving the prediction performance of the model.

[0206] Referring to Figure 11 , the embodiment of the present application provides a storm surge spatio-temporal prediction device based on a deep learning neural network, including:

[0207] A dataset construction unit for constructing a spatio-temporal variable dataset according to typhoon characteristics and spatio-temporal characteristics of the typhoon; wherein, the typhoon characteristics include typhoon paths, typhoon wind speeds, typhoon pressures, and storm surges corresponding to the typhoon paths;

[0208] An input parameter determination unit for determining the storm surge, the typhoon wind speed, and the typhoon pressure at each time step as input parameters according to the spatio-temporal variable dataset;

[0209] A neural network training unit for inputting the input parameters into a pre-constructed deep learning neural network for training, so that the deep learning neural network predicts the storm surge at the next time step of each time step as an output parameter;

[0210] A storm surge prediction unit for inputting the input parameters to be predicted into the trained deep learning neural network to obtain a predicted storm surge.

[0211] It can be understood that the content in the above method embodiments is applicable to the device embodiments. The functions specifically implemented in the device embodiments are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those in the above method embodiments.

[0212] In some alternative embodiments, the functions / operations mentioned in the block diagram may not occur in the order mentioned in the operation diagram. For example, depending on the functions / operations involved, two consecutive blocks shown may actually be executed substantially simultaneously or the blocks can sometimes be executed in the reverse order. In addition, the embodiments presented and described in the flowcharts of the present application are provided by way of example for the purpose of providing a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logical flows presented herein. Alternative embodiments are contemplated where the order of various operations is changed and where sub-operations described as part of a larger operation are executed independently.

[0213] Furthermore, although the present application has been described in the context of functional modules, it should be understood that, unless otherwise stated to the contrary, one or more of the functions and / or features may be integrated in a single physical device and / or software module, or one or more functions and / or features may be implemented in separate physical devices or software modules. It can also be understood that a detailed discussion of the actual implementation of each module is not necessary for understanding the present application. Rather, considering the attributes, functions, and internal relationships of the various functional modules in the devices disclosed herein, the actual implementation of the module will be understood within the ordinary skill of an engineer. Thus, those skilled in the art can implement the present application as set forth in the claims without undue experimentation. It can also be understood that the specific concepts disclosed are illustrative only and are not intended to limit the scope of the present application, which is determined by the full scope of the appended claims and their equivalents.

[0214] If the above-mentioned functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of this application. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs.

[0215] The logic and / or steps represented in the flowchart or described in other ways herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch and execute instructions from the instruction execution system, apparatus, or device), or in combination with these instruction execution systems, apparatus, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in combination with an instruction execution system, apparatus, or device.

[0216] More specific examples (non-exhaustive list) of computer-readable media include the following: electrical connection parts with one or more wirings (electronic devices), portable computer disk cartridges (magnetic devices), random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memories), fiber optic devices, and portable compact disc read-only memories (CDROM). Additionally, the computer-readable medium can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other media, then editing, interpreting, or processing it in other suitable ways as necessary, and then storing it in a computer memory.

[0217] It should be understood that various parts of the present application can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits with logic gate circuits for implementing logic functions on data signals, application specific integrated circuits with appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0218] In the description of this specification, the descriptions referring to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0219] Although the embodiments of the present application have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present application. The scope of the present application is defined by the claims and their equivalents.

[0220] The above has specifically described the preferred embodiments of the present application, but the present application is not limited to the described embodiments. Those skilled in the art can also make various equivalent deformations or substitutions without violating the spirit of the present application, and these equivalent deformations or substitutions are all included within the scope defined by the claims of the present application.

Claims

1. A spatio-temporal prediction method for storm surges based on a deep learning neural network, characterized in that, The method includes the following steps: Construct a spatio-temporal variable dataset according to typhoon characteristics and the spatio-temporal characteristics of typhoons; wherein, the typhoon characteristics include typhoon path, typhoon wind speed, typhoon air pressure, and the storm surge corresponding to the typhoon path; Determine the storm surge, the typhoon wind speed, and the typhoon air pressure at each time step as input parameters according to the spatio-temporal variable dataset; Input the input parameters into a pre-constructed deep learning neural network for training, so that the deep learning neural network predicts the storm surge at the next time step of each time step as the output parameter; Input the input parameters to be predicted into the trained deep learning neural network to obtain a predicted storm surge.

2. The storm surge spatio-temporal prediction method based on a deep learning neural network according to claim 1, wherein, The step of calculating the typhoon air pressure includes the following steps: Determine the air pressure at the calculation point in the typhoon area; The expression of the air pressure at the calculation point is: ； Among them, represents the air pressure at the calculation point; represents the peripheral air pressure of the typhoon; represents the central air pressure of the typhoon; is a parameter characterizing the typhoon system, calculated by an empirical formula and adjusted and corrected according to the radius of the maximum wind speed; is the distance from the calculation point to the typhoon center.

3. The storm surge spatio-temporal prediction method based on a deep learning neural network according to claim 1, wherein The step of calculating the typhoon wind speed includes the following steps: Determine the wind speed distribution of the typhoon according to the air pressure distribution of the typhoon and the gradient wind relationship; The expression of the wind speed distribution is: ； Among them, is the wind speed distribution, is the Coriolis parameter, ; represents the angular velocity of the Earth's rotation, represents the latitude, represents the air density; is the distance from the calculation point in the typhoon area to the typhoon center; Determine the wind speed generated by the movement of the typhoon as the moving wind speed; The expression of the moving wind speed is: ； Wherein, is the moving wind speed, and respectively represent the eastward component and the northward component of the moving speed of the typhoon center; Determine the wind speed of the symmetric typhoon model according to the wind speed distribution and the moving wind speed; The expression of the wind speed of the symmetric typhoon model is: ； Wherein, is the wind speed of the symmetric typhoon model, is the angle between the line connecting the calculation point and the typhoon center and the due east direction, is the angle between the gradient wind and the surface wind, and are correction factors; Determine the typhoon wind speed at each calculation point in the typhoon area according to the wind speed of the symmetric typhoon model; The expression of the typhoon wind speed is: ； Among them, is the typhoon wind speed; is an intermediate parameter, ; represents the background wind speed of the 10-meter wind speed component sourced from the ERA5 reanalysis dataset; and respectively represent the threshold distances selected and determined through verification experience.

4. The storm surge spatio-temporal prediction method based on a deep learning neural network according to claim 1, wherein The step of determining the storm surge corresponding to the typhoon path includes the following steps: Input the typhoon wind speed and the typhoon air pressure into the FVCOM model to obtain the storm surge in the format of triangular grid data; Use the inverse distance weighted interpolation method to convert the storm surge in the format of triangular grid data into the storm surge in the format of rectangular grid data as the storm surge corresponding to the typhoon path.

5. The storm surge spatio-temporal prediction method based on a deep learning neural network according to claim 1, characterized in that Before the step of inputting the input parameters into a pre-constructed deep learning neural network for training, the method further includes the following steps: Construct the deep learning neural network based on the vision transformer and incorporating the adaptive Fourier neural operator; On the basis of the architecture of the vision transformer, use the adaptive Fourier neural operator to replace and optimize the self-attention algorithm in the deep learning neural network to perform spatial mixing operations on spatial features and channel features.

6. The storm surge spatio-temporal prediction method based on a deep learning neural network according to claim 5, wherein The step of inputting the input parameters to be predicted into the trained deep learning neural network to obtain a predicted storm surge includes the following steps: Take the storm surge, the typhoon wind speed, the typhoon air pressure, and the corresponding time step at the corresponding time step in the input parameters as one channel respectively, and then divide the input parameters into several small data blocks; Use the encoder in the deep learning neural network to embed each small data block into a higher-dimensional space, and then apply position encoding to generate tokens corresponding to each small data block; Use the adaptive Fourier neural operator in the trained deep learning neural network to perform mixing processing on each token, and then determine the predicted storm surge according to the mixed tokens.

7. The storm surge spatio-temporal prediction method based on a deep learning neural network according to claim 6, characterized in that Performing hybrid processing on each of the tokens by using the adaptive Fourier neural operator in the trained deep learning neural network, and then determining the predicted storm surge according to the hybridized tokens, includes the following steps: Performing fast Fourier transform by using the adaptive Fourier neural operator to transform the small data blocks corresponding to each of the tokens into the frequency domain; Setting weights for the corresponding tokens by using the adaptive Fourier neural operator in the frequency domain through the diagonal weight matrix and soft threshold operation of the small data blocks; Performing non-linear mapping on each small data block according to the weights to improve the sparsity of each small data block; Repeating the token processing steps a set number of times to obtain the final tokens. The token processing steps include: for the spatial hybridization of each token, performing transformation by using a multi-layer MLP to obtain the hybridized tokens; converting the hybridized tokens back to the spatial domain through inverse Fourier transform, and then performing residual connection; for the channel hybridization of each token, performing transformation by using a single-layer MLP to achieve non-linear mapping of features; Using a decoder to embed the final tokens to the reconstruction time to obtain the predicted storm surge.

8. A storm surge spatio-temporal prediction device based on a deep learning neural network, characterized in that, The device includes: A dataset construction unit for constructing a spatio-temporal variable dataset according to typhoon characteristics and spatio-temporal characteristics of the typhoon; wherein, the typhoon characteristics include typhoon path, typhoon wind speed, typhoon air pressure, and the storm surge corresponding to the typhoon path; An input parameter determination unit for determining the storm surge, the typhoon wind speed, and the typhoon air pressure at each time step as input parameters according to the spatio-temporal variable dataset; A neural network training unit for inputting the input parameters into a pre-constructed deep learning neural network for training, so that the deep learning neural network predicts the storm surge at the next time step of each time step as an output parameter; A storm surge prediction unit for inputting the input parameters to be predicted into the trained deep learning neural network to obtain the predicted storm surge.

9. An electronic device, characterized in that, The electronic device includes a processor and a memory; The memory is used for storing programs; The processor executes the program to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a program, and the program is executed by the processor to implement the method according to any one of claims 1 to 7.

Images