Hurricane path prediction method and system based on recurrent neural network

Abstract

The application provides a hurricane path prediction method and system based on a recurrent neural network, comprising: carrying out grid processing on a delimited hurricane path area; calculating the moving distance and moving angle of each coordinate point as data enhancement; using a three-layer long short-term memory network as the main recurrent neural network of a prediction model; the main network is a long short-term memory model carrying a Dropout layer with a proportion coefficient of 0.1; the output layer is a fully connected layer activated by a hyperbolic tangent function, the model training uses a momentum-based gradient descent algorithm RMSProp, and the training target is the square error of the predicted longitude and latitude coordinates; the output layer of the prediction model is composed of a grid ID prediction layer and a coordinate correction layer to finally predict the hurricane coordinates. In the prediction stage, the inference prediction based on the recurrent neural network only depends on the forward propagation of the model, and it is not necessary to carry out associated search on the historical database, so that the performance of the hurricane path prediction is greatly improved.

Description

Technical Field

[0001] This invention relates to the field of information technology, and more specifically, to a hurricane path prediction method and system based on recurrent neural networks. Background Technology

[0002] Predicting the trajectory of moving targets is a popular topic in many research fields. Given the severity of the threats hurricanes pose to people's livelihoods, agriculture, and the economy, hurricane path prediction is an even more crucial issue within this area of ​​moving target trajectory prediction.

[0003] Current hurricane path prediction technologies are primarily based on numerical statistical methods. These methods typically predict hurricane paths using simulated hurricane paths obtained from simulation models. The advantage of statistical methods is their lightweight nature; their predictive power mainly relies on prior hurricane feature modeling. However, due to the lack of parameterization and learning capabilities of simulation models, they lack the ability to extract dynamic models that conform to big data trends from historical data. Another type of hurricane path prediction method is based on association analysis. This method mines frequent subsequences of hurricane movement paths from historical hurricane trajectory databases and then predicts the future path of the hurricane by searching for the most similar frequent subsequences to the predicted trajectory. Since sequence mining algorithms cannot be directly applied to trajectory data, it is usually necessary to convert the original hurricane trajectory data into hurricane spatiotemporal transaction data with coarser spatiotemporal granularity. This results in relatively coarse spatiotemporal granularity for path prediction based on association analysis. Furthermore, the additional time overhead from frequent sequence mining of spatiotemporal transaction databases is significant.

[0004] A patent document with publication number CN113642475A discloses a method for estimating the intensity of Atlantic hurricanes based on a convolutional neural network model. This method utilizes deep learning technology to analyze and learn from a large amount of geostationary satellite infrared cloud image data, automatically extracting complex features related to the intensity of tropical cyclones from their cloud system images, and then estimating the intensity. The model established in this invention uses a 114×114 image as the optimal input, combining four convolutional layers and two pooling layers, followed by three fully connected layers. The kernel size of the first three convolutional layers is 7×7, and the kernel size of the fourth convolutional layer is 3×3. Dropout terms are introduced before and after the convolutional layers and the fully connected layers. Finally, the estimated intensity is smoothed over an 18-hour time frame to obtain the final estimated hurricane intensity.

[0005] Therefore, a new technical solution is needed to improve the above-mentioned technical problems. Summary of the Invention

[0006] In view of the shortcomings of the prior art, the purpose of this invention is to provide a hurricane path prediction method and system based on recurrent neural networks.

[0007] According to the present invention, a hurricane path prediction method based on a recurrent neural network is provided, the method comprising the following steps:

[0008] Step S1: Grid the defined hurricane path area, mapping the latitude and longitude coordinates of the original hurricane trajectory data to grid IDs and grid coordinate correction coefficients;

[0009] Step S2: Calculate the movement distance and movement angle of each coordinate point as data augmentation, and obtain a 6-tuple sequence of trajectory point grid ID, coordinate correction coefficient, trajectory movement distance, movement angle, wind speed and central air pressure as input to the prediction model;

[0010] Step S3: A three-layer long short-term memory network is used as the backbone recurrent neural network of the prediction model. The input layer takes the preprocessed trajectory 6-tuple tensor as the data input; the backbone network is a long short-term memory model with a Dropout layer with a scaling factor of 0.1; the output layer is a fully connected layer activated by the hyperbolic tangent function. The model is trained using the momentum-based gradient descent algorithm RMSProp, and the training objective is to predict the squared error of latitude and longitude coordinates.

[0011] Step S4: The output layer of the prediction model consists of two parts: the grid ID prediction layer and the coordinate correction layer, which together form the final hurricane coordinate prediction.

[0012] Preferably, the gridding process for the defined hurricane path area in step S1 includes the following steps:

[0013] Step S1.1: Define the latitude and longitude range of the grid as 1°*1°, and number the 2D grid according to the Earth's plane;

[0014] Step S1.2: Map the latitude and longitude coordinates of the original trajectory points to their respective grid IDs, and then normalize the grid IDs by their maximum and minimum values. The grid ID values ​​are in the range of 0 to 1.

[0015] Preferably, the grid ID in step S1.2 is calculated as follows:

[0016] gridID = (lat-lat) min )*(lat max -lat min )+lon-lon min

[0017] gridID = (lat-lat) min )*(lon max -lon min )+lon-lon min

[0018] Preferably, the grid correction coefficient in step S1 is the latitude and longitude deviation of the coordinate point relative to the lower left corner of the grid. The latitude and longitude error of the hurricane coordinates caused by 1*1 gridding is the decimal part of the original coordinates. The grid correction coefficient is the decimal part of the original latitude and longitude coordinates. The normalized value of the grid ID where the hurricane trajectory point is located and the latitude and longitude correction value are used to replace the original latitude and longitude values.

[0019] Preferably, the long short-term memory network in step S3 is a time-recurrent neural network structure. A recurrent neural network is a chain-like structure with repeating modules. Each repeating result module is called a long short-term memory block. Each block is an intelligent network unit. In a block, there is a forget gate that determines whether the input of the block is important and will be remembered, an input gate that determines whether the input value in the block's memory affects the block, and an output gate that determines whether the block outputs information.

[0020] The present invention also provides a hurricane path prediction system based on a recurrent neural network, the system comprising the following modules:

[0021] Module M1: Performs gridding on the defined hurricane path area, mapping the latitude and longitude coordinates of the original hurricane trajectory data to grid IDs and grid coordinate correction coefficients;

[0022] Module M2: Calculates the distance and angle of movement for each coordinate point as data augmentation, and obtains a 6-tuple sequence of trajectory point grid ID, coordinate correction coefficient, trajectory movement distance, movement angle, wind speed, and central air pressure as input to the prediction model;

[0023] Module M3: Uses a three-layer Long Short-Term Memory (LSTM) network as the backbone recurrent neural network for the prediction model. The input layer takes a preprocessed trajectory 6-tuple tensor as the data input; the backbone network is an LSM model with a Dropout layer with a scaling factor of 0.1; the output layer is a fully connected layer activated by the hyperbolic tangent function. The model is trained using the momentum-based gradient descent algorithm RMSProp, and the training objective is to predict the squared error of latitude and longitude coordinates.

[0024] Module M4: The output layer of the prediction model consists of two parts: the grid ID prediction layer and the coordinate correction layer, which together form the final hurricane coordinate prediction.

[0025] Preferably, the gridding process for the defined hurricane path area in module M1 includes the following modules:

[0026] Module M1.1: Defines the latitude and longitude range of the grid as 1°*1°, and uses 2D grid numbering according to the Earth's plane;

[0027] Module M1.2: Maps the latitude and longitude coordinates of the original trajectory points to their respective grid IDs, and then normalizes the grid IDs by their maximum and minimum values. The grid ID values ​​range from 0 to 1.

[0028] Preferably, the grid ID in module M1.2 is calculated as follows:

[0029] gridID = (lat-lat) min )*(lat max -lat min )+lon-lon min

[0030] gridID = (lat-lat) min )*(lon max -lon min )+lon-lon min

[0031] Preferably, the grid correction coefficient in module M1 is the latitude and longitude deviation of the coordinate point relative to the lower left corner of the grid. The latitude and longitude error of the hurricane coordinates brought about by 1*1 gridding is the decimal part of the original coordinates. The grid correction coefficient is the decimal part of the original latitude and longitude coordinates. The normalized value of the grid ID where the hurricane trajectory point is located and the latitude and longitude correction value are used to replace the original latitude and longitude values.

[0032] Preferably, the long short-term memory network in module M3 is a time-recurrent neural network structure. A recurrent neural network is a chain-like structure with repeating modules. Each repeating result module is called a long short-term memory block. Each block is an intelligent network unit. In a block, there is a forget gate that determines whether the input of the block is important and will be remembered, an input gate that determines whether the input value in the block's memory affects the block, and an output gate that determines whether the block outputs information.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. Compared with hurricane motion simulation models, the hurricane path trajectory prediction based on recurrent neural networks in this invention can better fit the complex nonlinear path of the actual hurricane trajectory, thereby improving the accuracy and spatiotemporal granularity of hurricane path prediction.

[0035] 2. Compared with hurricane path prediction algorithms based on frequent pattern mining and matching, this invention provides finer granularity for hurricane path prediction and better accuracy during inference. In terms of performance, during the training phase, the convergence time of the deep learning model is significantly less than the time cost of mining the transaction spatiotemporal database.

[0036] 3. In the prediction phase, the recurrent neural network-based inference prediction relies only on the forward propagation of the model and does not require correlation search of historical databases, which greatly improves the performance of hurricane path prediction. Attached Figure Description

[0037] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0038] Figure 1 This is a schematic diagram of the process of the present invention;

[0039] Figure 2 This is a schematic diagram of the original hurricane trajectory points of this invention;

[0040] Figure 3 This is a schematic diagram of coordinate point gridding and coordinate correction coefficients in this invention;

[0041] Figure 4 This is a diagram of the long short-term memory recurrent neural network architecture of the present invention;

[0042] Figure 5 This is a schematic diagram of the long short-term memory block of the present invention;

[0043] Figure 6 This is a diagram illustrating the hurricane path prediction effect of the present invention. Detailed Implementation

[0044] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0045] Example 1:

[0046] According to the present invention, a hurricane path prediction method based on a recurrent neural network is provided, the method comprising the following steps:

[0047] Step S1: Grid the defined hurricane path area, mapping the latitude and longitude coordinates of the original hurricane trajectory data to grid IDs and grid coordinate correction coefficients;

[0048] The process of gridding the defined hurricane path area includes the following steps:

[0049] Step S1.1: Define the latitude and longitude range of the grid as 1°*1°, and number the 2D grid according to the Earth's plane;

[0050] Step S1.2: Map the latitude and longitude coordinates of the original trajectory points to their respective grid IDs, and then normalize the grid IDs by their maximum and minimum values. The grid ID values ​​are in the range of 0 to 1.

[0051] The grid ID is calculated as follows:

[0052] gridID = (lat-lat) min )*(lat max -lat min )+lon-lon min

[0053] gridID = (lat-lat) min )*(lon max -lon min )+lon-lon min

[0054] The grid correction factor is the latitude and longitude deviation of the coordinate point relative to the lower left corner of the grid. The latitude and longitude error of the hurricane coordinates brought about by 1*1 gridding is the decimal part of the original coordinates. The grid correction factor is the decimal part of the original latitude and longitude coordinates. The normalized value of the grid ID where the hurricane trajectory point is located and the latitude and longitude correction value are used to replace the original latitude and longitude values.

[0055] Step S2: Calculate the movement distance and movement angle of each coordinate point as data augmentation, and obtain a 6-tuple sequence of trajectory point grid ID, coordinate correction coefficient, trajectory movement distance, movement angle, wind speed and central air pressure as input to the prediction model.

[0056] Step S3: Use a three-layer long short-term memory network as the backbone recurrent neural network of the prediction model. The input layer takes the preprocessed trajectory 6-tuple tensor as the data input; the backbone network is a long short-term memory model with a Dropout layer with a scaling factor of 0.1; the output layer is a fully connected layer activated by the hyperbolic tangent function. The model is trained using the momentum-based gradient descent algorithm RMSProp, and the training objective is to predict the squared error of latitude and longitude coordinates.

[0057] Long Short-Term Memory (LSTM) networks are a type of recurrent neural network structure. Recurrent neural networks are chains of repeating modules. Each repeating module is called an LSM block, and each block is an intelligent network unit. Each block has a forget gate that determines whether the input to the block is important and will be remembered, an input gate that determines whether the input value in the block's memory affects the block, and an output gate that determines whether the block outputs information.

[0058] Step S4: The output layer of the prediction model consists of two parts: the grid ID prediction layer and the coordinate correction layer, which together form the final hurricane coordinate prediction.

[0059] Example 2:

[0060] Example 2 is a preferred embodiment of Example 1, and is used to illustrate the present invention in more detail.

[0061] This invention also provides a hurricane path prediction system based on a recurrent neural network, the system comprising the following modules:

[0062] Module M1: Performs gridding on the defined hurricane path area, mapping the latitude and longitude coordinates of the original hurricane trajectory data to grid IDs and grid coordinate correction coefficients;

[0063] The gridding process for the defined hurricane path area includes the following modules:

[0064] Module M1.1: Defines the latitude and longitude range of the grid as 1°*1°, and uses 2D grid numbering according to the Earth's plane;

[0065] Module M1.2: Maps the latitude and longitude coordinates of the original trajectory points to their respective grid IDs, and then normalizes the grid IDs by their maximum and minimum values. The grid ID values ​​range from 0 to 1.

[0066] The grid ID is calculated as follows:

[0067] gridID = (lat-lat) min )*(lat max -lat min )+lon-lon min

[0068] gridID = (lat-lat) min )*(lon max -lon min )+lon-lon min

[0069] The grid correction factor is the latitude and longitude deviation of the coordinate point relative to the lower left corner of the grid. The latitude and longitude error of the hurricane coordinates brought about by 1*1 gridding is the decimal part of the original coordinates. The grid correction factor is the decimal part of the original latitude and longitude coordinates. The normalized value of the grid ID where the hurricane trajectory point is located and the latitude and longitude correction value are used to replace the original latitude and longitude values.

[0070] Module M2: Calculates the movement distance and angle of each coordinate point as data augmentation, and obtains a 6-tuple sequence of trajectory point grid ID, coordinate correction coefficient, trajectory movement distance, movement angle, wind speed, and central air pressure as input to the prediction model.

[0071] Module M3: Uses a three-layer long short-term memory network as the backbone recurrent neural network for the prediction model. The input layer takes the preprocessed trajectory 6-tuple tensor as the data input; the backbone network is a long short-term memory model with a Dropout layer with a scaling factor of 0.1; the output layer is a fully connected layer activated by the hyperbolic tangent function. The model is trained using the momentum-based gradient descent algorithm RMSProp, and the training objective is to predict the squared error of latitude and longitude coordinates.

[0072] Long Short-Term Memory (LSTM) networks are a type of recurrent neural network structure. Recurrent neural networks are chains of repeating modules. Each repeating module is called an LSM block, and each block is an intelligent network unit. Each block has a forget gate that determines whether the input to the block is important and will be remembered, an input gate that determines whether the input value in the block's memory affects the block, and an output gate that determines whether the block outputs information.

[0073] Module M4: The output layer of the prediction model consists of two parts: the grid ID prediction layer and the coordinate correction layer, which together form the final hurricane coordinate prediction.

[0074] Example 3:

[0075] Example 3 is a preferred example of Example 1, and is used to illustrate the present invention in more detail.

[0076] To address the shortcomings of existing technologies, this invention provides a hurricane path prediction technique based on recurrent neural networks, aiming to improve the efficiency of hurricane feature characterization and the accuracy of hurricane path prediction. This technique uses recurrent neural networks to learn historical hurricane trajectory features and predicts the hurricane's path for the next 72 hours based on its current trajectory. The specific steps are as follows:

[0077] 1. Grid-based processing of the defined hurricane path area: Map the latitude and longitude coordinates of the original hurricane trajectory data to grid IDs and grid coordinate correction coefficients.

[0078] 2. Calculate the distance and angle of movement for each coordinate point as data augmentation. Obtain a 6-tuple sequence of trajectory point grid ID, coordinate correction coefficient, trajectory movement distance, movement angle, wind speed, and central air pressure as input to the prediction model.

[0079] 3. A three-layer Long Short-Term Memory (LSTM) network is used as the backbone recurrent neural network for the prediction model. The input layer uses preprocessed trajectory 6-tuple tensors as data input; the backbone network is an LSTM model with a Dropout layer having a scaling factor of 0.1; the output layer is a fully connected layer activated by the hyperbolic tangent function. The model is trained using the momentum-based gradient descent algorithm RMSProp, with the training objective being the squared error of the predicted latitude and longitude coordinates.

[0080] 4. The output layer of the prediction model consists of two parts: the grid ID prediction layer and the coordinate correction layer, which together form the final hurricane coordinate prediction.

[0081] The process of gridding the hurricane path area is as follows: The grid is defined with a latitude and longitude range of 1°*1°, and the grids are numbered in 2D according to the Earth's plane. The latitude and longitude coordinates of the original trajectory points are mapped to their respective grid IDs. Then, the grid IDs are normalized by performing maximum and minimum value normalization, ensuring that the grid ID values ​​are between 0 and 1. Normalized data can improve the generalization ability of the recurrent neural network prediction and accelerate the convergence speed of the training process.

[0082] The grid correction factor is the latitude and longitude deviation of the coordinate point relative to the lower left corner of the grid. Since the latitude and longitude error of the hurricane coordinates caused by 1*1 gridding is the decimal part of the original coordinates, the grid correction factor is the decimal part of the original latitude and longitude coordinates. The normalized value of the grid ID where the hurricane trajectory point is located and the latitude and longitude correction value are used to replace the original latitude and longitude values.

[0083] The Long Short-Term Memory (LSTM) Network described is a type of recurrent neural network. A recurrent neural network is a chain of repeating modules, each repeating module being called a Long Short-Term Memory block (LSTMCell). Each block is an intelligent network unit, containing a forget gate to determine whether the block's input is important and thus remembered, an input gate to determine whether the input values ​​in the block's memory affect the block, and an output gate to determine whether the block outputs information.

[0084] This embodiment, based on historical hurricane trajectory data from the western Pacific Ocean spanning 1920 to 2012, introduces the modeling and training process of a hurricane prediction model based on recurrent neural networks. For the original historical hurricane data trajectory point distribution, please refer to... Figure 2 .

[0085] First, the defined hurricane path area is gridded, mapping the latitude and longitude coordinates of the original hurricane trajectory data to grid IDs and grid coordinate correction coefficients. The grid ID is calculated as follows:

[0086] gridID = (lat-lat) min )*(lat max -lat min )+lon-lon min

[0087] gridID = (lat-lat) min )*(lon max -lon min )+lon-lon min

[0088] Where gridID is the grid ID of the trajectory point to be obtained;

[0089] lat and lon are the coordinates of the current trajectory point;

[0090] lat min lat max lon min These represent the minimum / maximum latitude and minimum longitude values ​​for all trajectory points appearing in the entire dataset, respectively.

[0091] Please refer to the diagram of coordinate correction factors and meshing. Figure 3 .

[0092] The following formulas are used to calculate the distance and angle of motion of each trajectory point relative to the previous trajectory point:

[0093] C = sin lat A *sin lat B *cos(lon A -lon B )+cis lat A *Cos lat B

[0094]

[0095] dy = sin(lon) A -lon B )*cos lat A

[0096] dx = cos lon A *sin lat A -sin lon A *cos lat A *cos(lonA -lon B )

[0097]

[0098] In the above formula, lat A lon A lat B lon B These are the latitude and longitude coordinates of the current trajectory point and the previous trajectory point, respectively; R is the Earth's radius constant; C represents the central angle of the two trajectory points relative to the reference ellipsoid; Distance is the distance between the two trajectory points; dx and dy represent the distances between the two trajectory points on the horizontal and vertical axes of the Earth's projection plane, respectively. Since the Earth's radius constant is directly canceled out here, there may be unit issues. If this is not easy to explain, you can directly substitute dx and dy into the formula for calculating Angle; Angle is the angle of motion of the two trajectory points.

[0099] The grid ID, wind speed, central air pressure, movement distance, and movement angle are normalized to their maximum and minimum values. The normalization is shown in the following formula:

[0100]

[0101] x norm This refers to the normalized value of the statistic, while the X value is the original value of the statistic.

[0102] Construct a three-layer LSTM backbone recurrent neural network. (See also...) Figure 4 and Figure 5 The diagram shows the LSTM block structure and the overall architecture of the three-layer LSTM backbone network. The output of the recurrent neural network is the predicted grid ID and coordinate correction values. The latitude and longitude coordinate prediction of the model is given by the following formula:

[0103] lon pred =lon min +gridID pred mod(lon max -lon min )+x pred

[0104]

[0105] gridID in the above formula pred , (x pred y pred These are the grid IDs predicted by the model, and the correction values ​​for longitude and latitude, respectively; lon pred lat predThese refer to the predicted values ​​of latitude and longitude coordinates, respectively. "mod" refers to the modulo operation, and is also denoted as %.

[0106] This embodiment uses a recurrent neural network with a prediction step size of 12, encapsulating the preprocessed trajectory point 6-tuples into tensors with a fixed step size (zero-padding for insufficient data). The momentum-based gradient descent algorithm RMSProp is used, with the training objective being the squared error of the predicted latitude and longitude coordinates. The learning rate is set to 0.1. The model converges within approximately 100 training epochs.

[0107] During the inference phase, using an input tensor with a fixed prediction step size (12), the model predicts the latitude and longitude coordinates of the hurricane within the step time (approximately 12*6=72 hours). For the actual prediction results of this embodiment, please refer to [link / reference]. Figure 6 .

[0108] LSTM (Long Short-Term Memory): A type of recurrent neural network. Recurrent neural networks are chains of repeating modules, each repeating module being called a long short-term memory block. Each block is an intelligent network unit.

[0109] Dropout: Randomly discarding data is a technique used to prevent overfitting in deep learning models during training. During deep learning model training, for any neural network unit, the model temporarily ignores it with a certain probability (input parameters, typically between 0.1 and 0.5).

[0110] RMSProp: Root Mean Square Propagation, a method for calculating elastic gradients in deep learning. During training, the gradient for optimizing the model's parameters is a weighted average of the squared terms of mini-batch stochastic gradients from the most recent time steps.

[0111] Those skilled in the art can understand this embodiment as a more specific description of Embodiment 1 and Embodiment 2.

[0112] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, enabling the system and its various devices, modules, and units to function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered both software modules implementing the method and structures within the hardware component.

[0113] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A hurricane path prediction method based on recurrent neural networks, characterized in that, The method includes the following steps: Step S1: Grid the defined hurricane path area, mapping the latitude and longitude coordinates of the original hurricane trajectory data to grid IDs and grid coordinate correction coefficients; Step S2: Calculate the movement distance and movement angle of each coordinate point as data augmentation, and obtain a 6-tuple sequence of trajectory point grid ID, coordinate correction coefficient, trajectory movement distance, movement angle, wind speed and central air pressure as input to the prediction model; Step S3: A three-layer long short-term memory network is used as the backbone recurrent neural network of the prediction model. The input layer takes the preprocessed trajectory 6-tuple tensor as the data input; the backbone network is a long short-term memory model with a Dropout layer with a scaling factor of 0.1; the output layer is a fully connected layer activated by the hyperbolic tangent function. The model is trained using the momentum-based gradient descent algorithm RMSProp, and the training objective is to predict the squared error of latitude and longitude coordinates. Step S4: The output layer of the prediction model consists of two parts: the grid ID prediction layer and the coordinate correction layer, which together form the final hurricane coordinate prediction. The grid correction coefficient in step S1 is the latitude and longitude deviation of the coordinate point relative to the lower left corner of the grid. The latitude and longitude error of the hurricane coordinates brought about by 1*1 gridding is the decimal part of the original coordinates. The grid correction coefficient is the decimal part of the original latitude and longitude coordinates. The normalized value of the grid ID where the hurricane trajectory point is located and the latitude and longitude correction value are used to replace the original latitude and longitude values.

2. The hurricane path prediction method based on recurrent neural networks according to claim 1, characterized in that, The gridding process for the defined hurricane path area in step S1 includes the following steps: Step S1.1: Define the latitude and longitude range of the grid as 1°*1°, and number the 2D grid according to the Earth's plane; Step S1.2: Map the latitude and longitude coordinates of the original trajectory points to their respective grid IDs, and then normalize the grid IDs by their maximum and minimum values. The grid ID values ​​are in the range of 0 to 1.

3. The hurricane path prediction method based on recurrent neural networks according to claim 2, characterized in that, The grid ID in step S1.2 is calculated as follows: in, Let be the grid ID of the desired trajectory point; , The coordinates of the current trajectory point; , These are the minimum and maximum latitude values ​​of all trajectory points appearing in the entire dataset, respectively. These represent the maximum and minimum longitude values ​​of all trajectory points appearing in the entire dataset, respectively.

4. The hurricane path prediction method based on recurrent neural networks according to claim 1, characterized in that, The Long Short-Term Memory (LSTM) network in step S3 is a time-recurrent neural network structure. A recurrent neural network is a chain-like structure with repeating modules. Each repeating result module is called a LSM block. Each block is an intelligent network unit. The block has a forget gate that determines whether the input of the block is important and will be remembered. It also has an input gate that determines whether the input value in the block's memory affects the block. Finally, it has an output gate that determines whether the block outputs information.

5. A hurricane path prediction system based on a recurrent neural network, characterized in that, The system includes the following modules: Module M1: Performs gridding on the defined hurricane path area, mapping the latitude and longitude coordinates of the original hurricane trajectory data to grid IDs and grid coordinate correction coefficients; Module M2: Calculates the distance and angle of movement for each coordinate point as data augmentation, and obtains a 6-tuple sequence of trajectory point grid ID, coordinate correction coefficient, trajectory movement distance, movement angle, wind speed, and central air pressure as input to the prediction model; Module M3: Uses a three-layer Long Short-Term Memory (LSTM) network as the backbone recurrent neural network for the prediction model. The input layer takes a preprocessed trajectory 6-tuple tensor as the data input; the backbone network is an LSM model with a Dropout layer with a scaling factor of 0.1; the output layer is a fully connected layer activated by the hyperbolic tangent function. The model is trained using the momentum-based gradient descent algorithm RMSProp, and the training objective is to predict the squared error of latitude and longitude coordinates. Module M4: The output layer of the prediction model consists of two parts: the grid ID prediction layer and the coordinate correction layer, which together form the final hurricane coordinate prediction. The grid correction coefficient in module M1 is the latitude and longitude deviation of the coordinate point relative to the lower left corner of the grid. The latitude and longitude error of the hurricane coordinates brought about by 1*1 gridding is the decimal part of the original coordinates. The grid correction coefficient is the decimal part of the original latitude and longitude coordinates. The normalized value of the grid ID where the hurricane trajectory point is located and the latitude and longitude correction value are used to replace the original latitude and longitude values.

6. The hurricane path prediction system based on a recurrent neural network according to claim 5, characterized in that, The gridding process for the defined hurricane path area in module M1 includes the following modules: Module M1.1: Defines the latitude and longitude range of the grid as 1°*1°, and uses 2D grid numbering according to the Earth's plane; Module M1.2: Maps the latitude and longitude coordinates of the original trajectory points to their respective grid IDs, and then normalizes the grid IDs by their maximum and minimum values. The grid ID values ​​range from 0 to 1.

7. The hurricane path prediction system based on a recurrent neural network according to claim 6, characterized in that, The grid ID in module M1.2 is calculated as follows: in, Let be the grid ID of the desired trajectory point; , The coordinates of the current trajectory point; , These are the minimum and maximum latitude values ​​of all trajectory points appearing in the entire dataset, respectively. These represent the maximum and minimum longitude values ​​of all trajectory points appearing in the entire dataset, respectively.

8. The hurricane path prediction system based on a recurrent neural network according to claim 5, characterized in that, The Long Short-Term Memory (LSTM) network in module M3 is a time-recurrent neural network structure. A recurrent neural network is a chain-like structure with repeating modules. Each repeating result module is called an LSM block. Each block is an intelligent network unit. In a block, there is a forget gate that determines whether the input of the block is important and will be remembered, an input gate that determines whether the input value in the block's memory affects the block, and an output gate that determines whether the block outputs information.

Images