A tropical cyclone track prediction method based on a space-time fusion network

Abstract

The application discloses a tropical cyclone track prediction method based on a space-time fusion network, and comprises the following steps: (1) obtaining an original data set, and constructing a three-dimensional time sequence structure of a tropical cyclone and surrounding environment data; (2) constructing a CaConvNeXt-GRU branch module to realize deep space-time modeling of the three-dimensional time sequence structure of the tropical cyclone; (3) constructing a convolution network branch module to extract two-dimensional time sequence features of the tropical cyclone; (4) constructing an overall DSTFN space-time fusion network model, and fusing the branch modules obtained in steps (2) and (3); (5) optimizing the space-time fusion network model; and (6) predicting two-dimensional and three-dimensional space-time features of the tropical cyclone; and the accuracy of the tropical cyclone track prediction method is higher than that of an existing deep learning-based tropical cyclone track prediction model.

Description

Technical Field

[0001] This invention relates to the field of cyclone trajectory prediction, and more specifically to a method for predicting the trajectory of tropical cyclones based on a spatiotemporal fusion network. Background Technology

[0002] With the rise of deep learning algorithms and the abundance of computing resources, meteorologists have begun to explore using deep learning algorithms to extract information from large climate datasets to solve the problem of tropical cyclone prediction. Deep learning models have advantages in processing large amounts of complex time-series data. On the one hand, deep learning models can extract implicit features from multivariate datasets, improving generalization ability; this can be achieved using convolutional neural networks (CNNs). Chen et al. used a CNN model to process the spatial relationships of tropical cyclone characteristics and formation. Liu et al. proposed an improved convolutional neural network model called ConvNeXt. ConvNeXt improves upon the standard ResNet model by borrowing the design ideas of the Swin Transformer. Experiments have shown that this model can achieve performance exceeding that of the Swin Transformer model. On the other hand, deep learning models are more effective as a method for solving time-series problems. Wang et al. used gated recurrent unit networks to extract temporal correlations from time-series datasets. To model the three-dimensional data of tropical cyclones—that is, to stack reanalysis data at different pressure levels to represent the three-dimensional structure of tropical cyclones—many spatiotemporal deep learning methods have been proposed. Fang et al. constructed the three-dimensional structure of tropical cyclones using reanalysis data and designed a Multi-ConvGRU model to extract features from this structure. Giffard-Roisin et al. proposed a deep learning model to fuse two-dimensional and three-dimensional data of tropical cyclones for tropical cyclone track prediction. This method uses a fully connected network to model the two-dimensional data of tropical cyclones to extract their nonlinear features. Simultaneously, a simple CNN model is used to model the three-dimensional data of tropical cyclones to capture their three-dimensional nonlinear features. However, this method cannot fully extract useful information from the two-dimensional data of tropical cyclones. Furthermore, due to the use of a large-scale reanalysis dataset, it is difficult to extract deep nonlinear spatial features using only a simple CNN model, nor can it process the complete three-dimensional structure of tropical cyclones. Additionally, this method does not consider the different effects of different isobaric surfaces on track prediction, and the use of only a CNN model fails to fully handle the temporal correlation of tropical cyclone time series. Xu et al. also explored a method for path prediction by fusing 2D and 3D data of tropical cyclones. Their paper proposed the AM-ConvGRU model, which introduces multiple convolutional modules and stacks residual channel attention (ReCA) modules based on residual learning, enabling the extraction of more complex spatial features from 3D tropical cyclone data. For 2D data, a generalized linear model was used. However, this paper suffers from insufficient training of the fusion model, simply fusing the features extracted from the two datasets through a fully connected network.Furthermore, this method only considers the influence of gravitational potential variables on the trajectory of tropical cyclones, without taking into account the influence of wind speed on the direction of movement of tropical cyclones. Summary of the Invention

[0003] Purpose of the invention: The purpose of this invention is to provide a tropical cyclone trajectory prediction method based on spatiotemporal fusion networks to solve the problem that traditional prediction methods are unable to fully capture the nonlinear spatiotemporal characteristics of the complex dynamic process of tropical cyclones, resulting in low prediction accuracy.

[0004] Technical solution: The tropical cyclone trajectory prediction method based on spatiotemporal fusion network described in this invention includes the following steps:

[0005] (1) Obtain the original dataset and construct a three-dimensional time-series structure of tropical cyclone and its surrounding environment data;

[0006] (2) Construct the CaConvNeXt-GRU branch module to realize in-depth spatiotemporal modeling of the three-dimensional temporal structure of tropical cyclones;

[0007] (3) Construct a convolutional network branch module to extract two-dimensional temporal features of tropical cyclones;

[0008] (4) Construct the overall DSTFN spatiotemporal fusion network model and fuse the branch modules obtained in steps (2) and (3);

[0009] (5) Optimize the spatiotemporal fusion network model;

[0010] (6) Predict the two-dimensional and three-dimensional spatiotemporal characteristics of tropical cyclones.

[0011] Further, step (1) is as follows: using the gravitational potential variables and wind speed data in the ERA 5 dataset, combined with the trajectory data in the IBTRCS dataset, a three-dimensional time-series structure of tropical cyclones and their surrounding environment data is constructed, and used as input data for the CaConvNeXt-GRU branch module; wherein, the four pressure levels include: 1000, 750, 500 and 250 hPa; the wind speed includes two components, U and V; wherein the trajectory data includes: the longitude and latitude of the tropical cyclone center location.

[0012] Furthermore, in step (2), CaConvNeXt-GRU is composed of a spatial feature extraction module, a downsampling module, a convolutional block attention module, and a stacked GRU model, including the following steps:

[0013] (21) Spatial Feature Extraction Module: First, the input data undergoes initial downsampling through a convolutional layer and layer normalization; the kernel size is 4×4 and the stride is 4; then it is input into a convolutional layer with a kernel size of 7×7 to perform dimensionality upsampling; second, the activation function GELU is used to introduce nonlinearity and enhance the network's expressive power; third, after dimensionality reduction through a convolutional layer with a kernel size of 1×1, it is input into a random depth regularization layer to prevent overfitting by randomly discarding some layers in the network; finally, residual connections are used to input the processed features and the original input into the downsampling module together; the spatial feature extraction module performs 3 stacking operations.

[0014] (22) Downsampling module: Uses convolution and normalization to reduce feature size;

[0015] (23) Convolutional Block Attention Module: Includes two modules: channel attention and spatial attention; the formula for channel attention is as follows:

[0016]

[0017] Where σ represents the Sigmoid function, W0∈R C / r×C W1∈R C / C×r C represents the number of channels in the feature map, and r represents the reduction rate. The output data of the two max pooling AvgPool(F) and average pooling AvgPool(F) share the MLP weights W0 and W1;

[0018] The formula for spatial attention is as follows:

[0019]

[0020] Where σ represents the Sigmoid function, f 7×7 This indicates a convolutional layer with a kernel size of 7×7;

[0021] (24) The spatial feature extraction modules are stacked 9, 3, and 3 times respectively. Before each stacking of the spatial feature extraction modules, they are processed by the convolutional block attention module and a separate downsampling module. After global average pooling and layer normalization, the feature data is converted into global pooling feature vectors. The data is then input into the GRU model to mine the temporal features in the relevant spatial range. Finally, the tropical cyclone prediction results after t+24 hours are obtained through a fully connected network.

[0022] Furthermore, step (3) is as follows: a one-dimensional convolutional neural network and a fully connected neural network are used to stack the past displacement of the tropical cyclone with the tropical cyclone trajectory features over time, and this is used as the input data for the branch modules of the convolutional network to model its temporal dependence, finally obtaining the tropical cyclone prediction result after t+24 hours; the formula is as follows:

[0023] Y=σ(conv1(σ(conv2(xt))))*w+b

[0024] Where conv1 and conv2 are two convolutional networks, w and b are the weights and biases, and σ is the GELU activation function, as shown in the following formula:

[0025]

[0026]

[0027] Where X ~ N(0,1), It is the cumulative distribution function of the standard normal distribution.

[0028] Furthermore, step (4) is as follows: a phased training strategy is adopted: first, each branch module is pre-trained independently to obtain the optimal parameters; then, the branch parameters are fixed and only the fusion module is trained; finally, each weight is re-optimized to retrain the entire DSTFN model to fine-tune all parameters of the network.

[0029] Furthermore, in step (5), the model is evaluated using the Haversine formula, as follows:

[0030]

[0031] Where N is the total number of samples, D i is the Haversine distance of the i-th sample, which is calculated using the great circle distance to determine the error distance between the ground truth point and the predicted point on the sphere; Li(θ) is the sum of the absolute values ​​of the weights, and α is the regularization coefficient, which controls the contribution of the regularization term to the loss function;

[0032] Cross-validation is used as the training and validation strategy. Mean Squared Error (MSE) and Mean Absolute Error (MAE) are chosen as the criteria for the detection model's score. MSE and MAE can be expressed as follows:

[0033]

[0034]

[0035] The subscript of the output y has three representations: train, label, and test, which represent the output of the training process, the sample label, and the test output, respectively.

[0036] The tropical cyclone trajectory prediction system based on spatiotemporal fusion network of the present invention includes:

[0037] Acquisition module: Used to acquire raw datasets and construct a three-dimensional time-series structure of tropical cyclone and surrounding environmental data;

[0038] CaConvNeXt-GRU branch module: used to construct CaConvNeXt-GRU branch modules to achieve in-depth spatiotemporal modeling of the three-dimensional temporal structure of tropical cyclones;

[0039] Convolutional network branch module: Used to construct convolutional network branch modules to extract two-dimensional temporal features of tropical cyclones;

[0040] Spatiotemporal fusion network module: used to construct the overall spatiotemporal fusion network model and fuse the branch modules obtained in steps (2) and (3);

[0041] Optimization module: Used to optimize the spatiotemporal fusion network model;

[0042] Prediction module: Used to predict the two-dimensional and three-dimensional spatiotemporal characteristics of tropical cyclones.

[0043] An electronic device according to the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements any of the methods for predicting tropical cyclone trajectories based on spatiotemporal fusion networks.

[0044] The present invention provides a storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements any one of the tropical cyclone trajectory prediction methods based on a spatiotemporal fusion network.

[0045] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: The CaConvNeXt-GRU model can effectively achieve deep spatiotemporal modeling of the three-dimensional temporal structure of tropical cyclones, and can automatically select high-response isobaric surfaces to emphasize information more critical for predicting the future direction of tropical cyclones. Through a phased training strategy, the advantages of different modules of the model are leveraged, and effective fusion of multi-source heterogeneous data is achieved. The accuracy of the tropical cyclone track prediction method proposed in this invention is superior to existing deep learning-based tropical cyclone track prediction models. Attached Figure Description

[0046] Figure 1This is a flowchart of the present invention;

[0047] Figure 2 This is a structural diagram of the CaConvNeXt-GRU of the present invention;

[0048] Figure 3 This is a structural diagram of the spatial feature extraction module of the present invention;

[0049] Figure 4 This is a structural diagram of the downsampling module of the present invention;

[0050] Figure 5 This is a structural diagram of the DSTFN model of the present invention;

[0051] Figure 6 This is a three-dimensional time-series structure diagram of a tropical cyclone at an atmospheric pressure level according to the present invention;

[0052] Figure 7 This is the three-dimensional time-series structure of tropical cyclone and its surrounding environment data according to the present invention. Detailed Implementation

[0053] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0054] like Figure 1 As shown, this embodiment of the invention provides a method for predicting the trajectory of tropical cyclones based on a spatiotemporal fusion network, including the following steps:

[0055] (1) Obtain the original dataset and construct a three-dimensional time-series structure of tropical cyclone and its surrounding environment data; use the gravitational potential variable and wind speed data in the ERA5 dataset and combine them with the trajectory data of IBTRAS in the western Pacific from 2000 to 2022 to construct a three-dimensional time-series structure of tropical cyclone and its surrounding environment data, and use it as input data for the CaConvNeXt-GRU branch module; the four pressure levels include 1000, 750, 500 and 250 hPa; the wind speed includes two components, U and V; the trajectory data includes the longitude and latitude of the center of the tropical cyclone.

[0056] (2) Figure 2 As shown, the CaConvNeXt-GRU branch module is constructed to achieve deep spatiotemporal modeling of the three-dimensional temporal structure of tropical cyclones. CaConvNeXt-GRU consists of a spatial feature extraction module, a downsampling module, a convolutional block attention module, and a stacked GRU model, and includes the following steps:

[0057] (21) such as Figure 3As shown, the spatial feature extraction module performs the following steps: First, the input data undergoes initial downsampling through a convolutional layer and layer normalization; the kernel size is 4×4 with a stride of 4. Then, it is input into a convolutional layer with a kernel size of 7×7 for dimensionality upsampling. Second, the GELU activation function is used to introduce non-linearity and enhance the network's expressive power. Third, after dimensionality reduction through a convolutional layer with a kernel size of 1×1, the data is input into a random depth regularization layer, which randomly discards some layers to prevent overfitting. Finally, residual connections are used to input the processed features along with the original input into the downsampling module. The spatial feature extraction module involves three stacking operations.

[0058] (22) such as Figure 4 As shown, the downsampling module reduces feature size by using convolution and normalization.

[0059] (23) Convolutional Block Attention Module: Includes two modules: channel attention and spatial attention; the formula for channel attention is as follows:

[0060]

[0061] Where σ represents the Sigmoid function, W0∈R C / r×C W1∈R C / C×r C represents the number of channels in the feature map, and r represents the reduction rate. The output data of the two max pooling AvgPool(F) and average pooling AvgPool(F) share the MLP weights W0 and W1;

[0062] The formula for spatial attention is as follows:

[0063]

[0064] Where σ represents the Sigmoid function, f 7×7 This indicates a convolutional layer with a kernel size of 7×7;

[0065] (24) The spatial feature extraction modules are stacked 9, 3, and 3 times respectively. Before each stacking of the spatial feature extraction modules, they are processed by the convolutional block attention module and a separate downsampling module. After global average pooling and layer normalization, the feature data is converted into global pooling feature vectors. The data is then input into the GRU model to mine the temporal features in the relevant spatial range. Finally, the tropical cyclone prediction results after t+24 hours are obtained through a fully connected network.

[0066] (3) Constructing a convolutional network branch module to extract two-dimensional temporal features of tropical cyclones; specifically as follows: using a one-dimensional convolutional neural network and a fully connected neural network to stack the past displacement of tropical cyclones with their trajectory features over time, and using this as input data for the convolutional network branch module, modeling their temporal dependence, and finally obtaining the tropical cyclone prediction results after t+24 hours; the formula is as follows:

[0067] Y=σ(conv1(σ(conv2(xt))))*w+b

[0068] Where conv1 and conv2 are two convolutional networks, w and b are the weights and biases, and σ is the GELU activation function, as shown in the following formula:

[0069]

[0070]

[0071] Where X ~ N(0,1), It is the cumulative distribution function of the standard normal distribution.

[0072] (4) Construct the overall DSTFN spatiotemporal fusion network model and fuse the branch modules obtained in steps (2) and (3); specifically as follows: adopt a phased training strategy: first, pre-train each branch module independently to obtain the optimal parameters; then, fix the branch parameters and train only the fusion module; finally, allow re-optimization of each weight to retrain the entire DSTFN model to fine-tune all network parameters. Specifically as follows:

[0073] like Figure 5 As shown, firstly, the CaConvNeXt-GRU branch module is pre-trained; the input data for this module is the three-dimensional temporal structure of the tropical cyclone and its surrounding environment at the current time t and at the past 6, 12, and 18 times, i.e., X. t X t-6 X t-12 and X t-18 Each input data point has a size of dim × h × w, where dim represents the channel size of the input data, h represents the feature height, and w represents the feature width. The channel size array dim is set to (96, 192, 384, 768). The output data is Y. t+24The process involves several steps. First, the model is pre-trained. The first step is to obtain the tropical cyclone path prediction result for the branch module 24 hours from the current time. The optimal result is then saved. Next, a convolutional network branch module is pre-trained. The input data for this module consists of time-series data constructed by stacking three past displacements (from t-18h to t-12h, from t-12h to t-6h, and from t-6h to t) with tropical cyclone trajectory features. This results in the branch module's predicted tropical cyclone path 24 hours from the current time. The optimal result is then saved. Finally, the fusion module is trained with fixed branch module parameters. Each branch module is pre-trained independently to obtain... After optimizing the parameters, a fusion module is introduced to jointly model the 2D and 3D data of tropical cyclones using the best-performing modules, learning the correlations between heterogeneous data. This fusion module effectively combines the features extracted by different branch modules, improving the overall model's performance and generalization ability. Finally, the tropical cyclone track prediction result 24 hours after the current time is obtained from the jointly trained module. After achieving the optimal result, the model is saved. Finally, overall training is performed. After joint training, the entire DSTFN model is retrained based on the 2D and 3D tropical cyclone data, allowing for re-optimization of each weight to fine-tune all network parameters. Ultimately, the tropical cyclone trajectory prediction result 24 hours after overall training is obtained. Through the above optimization strategies and module design, effective fusion of various modules can be achieved while efficiently extracting multi-source features.

[0074] (5) Optimize the spatiotemporal fusion network model; evaluate the model using the Haversine formula, as follows:

[0075]

[0076] Where N is the total number of samples, D i L is the Haversine distance of the i-th sample, used to calculate the error distance between the actual ground point and the predicted point on the sphere using the great circle distance; i (θ) is the sum of the absolute values ​​of the weights, and α is the regularization coefficient, which controls the contribution of the regularization term to the loss function.

[0077] Cross-validation is used as the training and validation strategy. Mean Squared Error (MSE) and Mean Absolute Error (MAE) are chosen as the criteria for the detection model's score. MSE and MAE can be expressed as follows:

[0078]

[0079]

[0080] The subscript of the output y has three representations: train, label, and test, which represent the output of the training process, the sample label, and the test output, respectively.

[0081] (6) Predict the two-dimensional and three-dimensional spatiotemporal characteristics of tropical cyclones. The implementation process is as follows:

[0082] After all models were built, a three-dimensional and a two-dimensional temporal structure of tropical cyclones and their surrounding environment were constructed and input into the models for training. First, the two-dimensional temporal structure of tropical cyclones was constructed. Tropical cyclone trajectory data for the Northwest Pacific region were extracted from the IBTRASC dataset of the Western Pacific from 2000 to 2022. The positions of the tropical cyclone centers at different time points in the trajectory dataset were analyzed, and the difference between the center positions at adjacent time points, i.e., the displacement value (δlong), was calculated. Δt ,δlong Δt ), where Δt is a multiple of 6, and the time points include the current time t and the next 6, 12, and 18 hours (th, t-6h, t-12h, t-18h). During model training, feature data related to tropical cyclone trajectories obtained from the IBTRASC dataset (such as longitude, latitude, maximum sustained wind speed, air pressure, and distance from the center to land) are combined with the current displacement value (between time point t and time point t-6) and past displacement values ​​(between time point t-6 and time point t-12, and between time point t-12 and time point t-18) as input features for the convolutional network branch modules. This design aims to simultaneously capture the short-term and long-term displacement patterns of tropical cyclones and consider environmental features related to changes in tropical cyclone trajectories to better predict the trajectory path of tropical cyclones in the next 24 hours.

[0083] Then, a three-dimensional temporal structure of tropical cyclones is constructed. The path of a tropical cyclone is mainly influenced by internal dynamic factors and external environmental factors. Gravitational potential variables mainly describe the internal energy structure and dynamics of a tropical cyclone, while wind speed (including U and V velocity components) describes the intensity of the tropical cyclone and directly observes the atmospheric flow of the tropical cyclone. Therefore, the gravitational potential variables and wind speeds in the ECMWF atmospheric reanalysis dataset EAR 5 can be used to construct a three-dimensional temporal structure of tropical cyclones and their surrounding environment. First, the Earth is divided into a sphere composed of 1°×1° longitude and latitude surfaces. Then, a 25°×25° surface is intercepted centered on the longitude and latitude of the tropical cyclone, as shown below. Figure 6As shown, this surface incorporates atmospheric reanalysis data from within the tropical cyclone and information from its external environment. Next, at four atmospheric pressure levels (250, 500, 750, and 1000 hPa), the gravitational potential variable and the U and V velocity components of wind speed were extracted for each time step near the tropical cyclone. Based on the atmospheric reanalysis data, a three-dimensional temporal structure of the tropical cyclone and its surrounding environment was constructed along four dimensions: longitude, latitude, variables, and time. Figure 7 As shown, Z, U, and V represent the gravitational potential variable and the wind speed, respectively.

[0084] Based on two-dimensional and three-dimensional time-series data, pre-training, joint training, and overall training are performed. The final output is the tropical cyclone track prediction result for the next 24 hours. The overall DSTFN model can be represented as:

[0085] Q = F2(F1(f1(X) t ,X t-6 ,X t-12 ,X t-18 ),f2(Y t ,Y t-6 ,Y t-12 ,Y t-18 ),

[0086] f3(Z t Z t-6 Z t-12 Z t-18 )))

[0087] Where Q is the final prediction result, F2 is the overall training method, F1 is the joint training and fusion module method, f1 and f2 are the CaConvNeXt-GRU branch module methods, and X t -X t-18 and Y t -Y t-18 These are the three-dimensional temporal structures of tropical cyclones and their surrounding environment based on wind speed and gravitational potential variables, respectively; f3 is the convolutional network branch module method; and Z... t -Z t-18 It is a time series composed of tropical cyclone displacement and meteorological factors related to its trajectory.

[0088] The tropical cyclone trajectory prediction system based on spatiotemporal fusion network of the present invention includes:

[0089] Acquisition module: Used to acquire raw datasets and construct a three-dimensional time-series structure of tropical cyclone and surrounding environmental data;

[0090] CaConvNeXt-GRU branch module: used to construct CaConvNeXt-GRU branch modules to achieve in-depth spatiotemporal modeling of the three-dimensional temporal structure of tropical cyclones;

[0091] Convolutional network branch module: Used to construct convolutional network branch modules to extract two-dimensional temporal features of tropical cyclones;

[0092] Spatiotemporal fusion network module: used to construct the overall spatiotemporal fusion network model and fuse the branch modules obtained in steps (2) and (3);

[0093] Optimization module: Used to optimize the spatiotemporal fusion network model;

[0094] Prediction module: Used to predict the two-dimensional and three-dimensional spatiotemporal characteristics of tropical cyclones.

[0095] An electronic device according to the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements any of the methods for predicting tropical cyclone trajectories based on spatiotemporal fusion networks.

[0096] The present invention provides a storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements any one of the tropical cyclone trajectory prediction methods based on a spatiotemporal fusion network.

Claims

1. A method for predicting the trajectory of tropical cyclones based on a spatiotemporal fusion network, characterized in that, Includes the following steps: (1) Obtain the original dataset and construct a three-dimensional time-series structure of tropical cyclone and its surrounding environment data; (2) Constructing the CaConvNeXt-GRU branch module to achieve in-depth spatiotemporal modeling of the three-dimensional temporal structure of tropical cyclones; CaConvNeXt-GRU consists of a spatial feature extraction module, a downsampling module, a convolutional block attention module, and a stacked GRU model, including the following steps: (21) Spatial feature extraction module: First, the input data is initially downsampled through a convolutional layer and layer normalization; the kernel size is 4×4 and the stride is 4; then it is input into a convolutional layer with a kernel size of 7×7 to perform dimensionality upscaling; second, the activation function GELU is used to introduce nonlinearity and enhance the network's expressive power; third, after dimensionality reduction through a convolutional layer with a kernel size of 1×1, it is input into a random depth regularization layer to prevent overfitting by randomly discarding some layers in the network; finally, residual connections are used to input the processed features and the original input into the downsampling module together; the spatial feature extraction module is stacked 3 times. (22) Downsampling module: Uses convolution and normalization to reduce feature size; (23) Convolutional block attention module: includes two modules: channel attention and spatial attention; among which, the formula for channel attention is as follows: ; in, This represents the Sigmoid function. , , For the channels of the feature map, To reduce the rate, two max pooling methods are used. and average pooling The output data is shared. weight and ; The formula for spatial attention is as follows: ； in, This represents the Sigmoid function. Indicates the kernel size as Convolutional layers; (24) The spatial feature extraction modules are stacked 9, 3, and 3 times respectively. Before each stacking of the spatial feature extraction modules, they are processed by the convolutional block attention module and a separate downsampling module. After global average pooling and layer normalization, the feature data is converted into global pooling feature vectors. The data is then input into the GRU model to mine the temporal features in the relevant spatial range. Finally, the tropical cyclone prediction results after t+24 hours are obtained through a fully connected network. (3) Constructing a convolutional network branch module to extract two-dimensional temporal features of tropical cyclones; specifically as follows: using a one-dimensional convolutional neural network and a fully connected neural network to stack the past displacement of tropical cyclones with their trajectory features over time, and using this as input data for the convolutional network branch module, modeling its temporal dependency, and finally obtaining Tropical cyclone forecast results 24 hours later; the formula is as follows: ； in, , For two convolutional networks, For weights and biases, The GELU activation function is shown in the following formula: ； ； in, , It is the cumulative distribution function of the standard normal distribution; (4) Construct the overall DSTFN spatiotemporal fusion network model and fuse the branch modules obtained in steps (2) and (3); specifically as follows: adopt a phased training strategy: first, pre-train each branch module independently to obtain the optimal parameters; then, fix the branch parameters and train only the fusion module; finally, allow re-optimization of each weight to retrain the entire DSTFN model to fine-tune all network parameters. (5) Optimize the spatiotemporal fusion network model; evaluate the model using the Haversine formula, as follows: ； Where N is the total number of samples, D i It is the Haversine distance of the i-th sample, and the error distance between the ground truth point and the predicted point on the sphere is calculated using the great circle distance; It is the sum of the absolute values ​​of the weights, and α is the regularization coefficient, which controls the contribution of the regularization term to the loss function; Cross-validation is used as the training and validation strategy. Mean Squared Error (MSE) and Mean Absolute Error (MAE) are chosen as the criteria for the detection model's score. MSE and MAE can be expressed as follows: ； ； The subscript of the output y has three representations: train, label, and test, which represent the output of the training process, the sample label, and the test output, respectively. (6) Predict the two-dimensional and three-dimensional spatiotemporal characteristics of tropical cyclones.

2. The method for predicting tropical cyclone trajectories based on a spatiotemporal fusion network according to claim 1, characterized in that, The specific steps (1) are as follows: using the gravitational potential variables and wind speed data in the ERA 5 dataset, and combining them with the trajectory data in the IBTACS dataset, a three-dimensional time-series structure of tropical cyclones and their surrounding environment data is constructed, and used as input data for the CaConvNeXt-GRU branch module; among them, the four pressure levels include: 1000, 750, 500 and 250 hPa; the wind speed includes two components, U and V; and the trajectory data includes: the longitude and latitude of the center of the tropical cyclone.

3. A tropical cyclone trajectory prediction system based on a spatiotemporal fusion network, employing the method described in any one of claims 1-2, characterized in that, include: Acquisition module: Used to acquire raw datasets and construct a three-dimensional time-series structure of tropical cyclone and surrounding environmental data; CaConvNeXt-GRU branch module: used to construct CaConvNeXt-GRU branch modules to achieve in-depth spatiotemporal modeling of the three-dimensional temporal structure of tropical cyclones; Convolutional network branch module: Used to construct convolutional network branch modules to extract two-dimensional temporal features of tropical cyclones; Spatiotemporal fusion network module: used to construct the overall spatiotemporal fusion network model and fuse the branch modules obtained in steps (2) and (3); Optimization module: Used to optimize the spatiotemporal fusion network model; Prediction module: Used to predict the two-dimensional and three-dimensional spatiotemporal characteristics of tropical cyclones.

4. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is loaded into the processor, it implements a method for predicting the trajectory of a tropical cyclone based on a spatiotemporal fusion network according to any one of claims 1-2.

5. A storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements a method for predicting the trajectory of a tropical cyclone based on a spatiotemporal fusion network according to any one of claims 1-2.

Images