An artificial intelligence driven typhoon path dynamic tracking method and system

Abstract

The application provides an artificial intelligence driven typhoon path dynamic tracking method and system, the system comprises a running environment, a data control program and a typhoon positioning program, the data control program is responsible for managing data flow input and output, including initial coordinate processing, meteorological data preprocessing and data post-processing process, the path tracking coordinates are updated through cyclic iteration, and finally complete typhoon tracking data is output, and the typhoon positioning program adopts a three-layer progressive processing architecture, including dynamic window tracking, multi-scale feature fusion network reasoning and path coordinate generation.The application has the beneficial effects that based on the deep learning multi-scale feature fusion network, the typhoon path can be tracked accurately and quickly by using numerical weather prediction data, the running environment is compatible with the running of GPU and CPU environments, and the running efficiency is much higher than that of traditional positioning algorithms.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of meteorological prediction and artificial intelligence, and particularly relates to an artificial intelligence driven typhoon path dynamic tracking method and system. BACKGROUND

[0002] Typhoon path tracking, as a core task of meteorological prediction, has long been highly concerned by the academic and business communities at home and abroad. Traditional methods mainly rely on numerical weather prediction models (such as WRF, ECMWF) and statistical models to achieve prediction by solving atmospheric dynamics equations or fitting historical data. In recent years, with the advancement of international cooperation projects and technology iteration, the accuracy of typhoon path prediction has gradually improved. For example, the Typhoon Landing Forecast Demonstration Project (TLFDP) led by the World Meteorological Organization (WMO) has developed 14 new technologies for typhoon path, intensity and precipitation prediction by integrating the technological achievements of multiple national institutions, significantly improving the reliability of the forecast. In terms of specific technology, the vortex tracking technology (GVT) proposed by the American GFDL (Geophysical Fluid Dynamics Laboratory) has become one of the representative methods. This technology effectively corrects the "zigzag" swing problem that traditional methods (such as the minimum sea level pressure method and the 500 hPa vorticity maximum method) are prone to in path prediction by locating the core area of the typhoon vortex. Studies have shown that the application of GVT in the Zhejiang regional model makes the typhoon path prediction closer to the actual observation, especially in the control of the positioning error of the lowest pressure in the typhoon center.

[0003] Despite the progress made by traditional methods in business applications, their limitations have become increasingly apparent:

[0004] 1. Lack of timeliness: Traditional methods take relatively long time to generate typhoon path tracking, and consume a lot of resources.

[0005] 2. Limited ability to capture nonlinear features: The structural and property changes and path mutations in the dynamic evolution of typhoons often lead to positioning errors. SUMMARY

[0006] Therefore, the present application aims to provide an artificial intelligence driven typhoon path dynamic tracking method and system to solve the problems of lack of timeliness and limited ability to capture nonlinear features in the prior art.

[0007] To achieve the above-mentioned purpose, the technical solution of the present application is as follows:

[0008] In a first aspect, the present application provides an artificial intelligence driven typhoon path dynamic tracking method, comprising the following steps:

[0009] S1, training a multi-scale feature fusion network MSF-Net model;

[0010] The MSF-Net model comprises the following modules: a primary feature encoding module for extracting texture features of an input meteorological field step by step through convolution operation; a deep feature enhancement module for introducing Gaussian noise in a convolution layer and adding input features and output features through cross-layer connection; a global context fusion module for integrating multi-scale features by using a convolution kernel and generating a global feature vector; a feature compression module for compressing the global feature vector through a fully connected network; and a relative position index prediction module for outputting a relative position index of a typhoon center.

[0011] S2, based on the original numerical weather prediction data, the meteorological data processing layer performs meteorological data fusion and grid resampling, and outputs processed numerical weather prediction data;

[0012] S3, based on the initial information of the typhoon, the initial coordinate processing layer determines the grid index and boundary range of the typhoon center; and the geographic coordinate mapping layer establishes a bidirectional conversion relationship between the geographic coordinates and the grid index;

[0013] S4, the typhoon path tracking coordinates are updated iteratively;

[0014] S5, outputting the typhoon path tracking data, and performing Gaussian smoothing and quality control processing on the typhoon path tracking data;

[0015] In step S3, based on the initial information of the typhoon, the initial coordinate processing layer determines the grid index and boundary range of the typhoon center, comprising:

[0016] Based on the initial position of the typhoon, the initial coordinate processing layer generates the initial latitude and longitude index and the initial latitude and longitude boundary index of the typhoon center;

[0017] In step S4, the typhoon path tracking coordinates are updated iteratively, comprising:

[0018] S41, based on the processed numerical weather prediction data, the latitude and longitude index and the latitude and longitude boundary index of the typhoon center, the sliding window extraction layer extracts a local meteorological field;

[0019] S42, based on the extracted local meteorological field, the multi-scale feature fusion network MSF-Net inference layer predicts the relative position offset of the typhoon center;

[0020] S43, based on the predicted relative position of the typhoon center, the path coordinate generation layer updates the latitude and longitude index and the latitude and longitude boundary index of the typhoon center;

[0021] S44, the updated latitude and longitude index and latitude and longitude boundary index of the typhoon center are fed back to the sliding window extraction layer for the next round of circulation.

[0022] Further, in step S1, the multi-scale feature fusion network MSF-Net model is trained, including:

[0023] S11, divide the typhoon path data into a training set, a validation set and a test set;

[0024] S12, use the training set, the validation set and the test set to realize learning, optimization and evaluation of the multi-scale feature fusion network MSF-Net model;

[0025] S13, save the multi-scale feature fusion network MSF-Net model.

[0026] Further, in step S2, the meteorological data processing layer performs meteorological data fusion and grid resampling, including:

[0027] S21, meteorological data fusion: integrate meteorological parameter channels covering near-surface meteorological element fields and upper air meteorological element fields;

[0028] S22, grid resampling processing: using a bilinear interpolation algorithm, resample the meteorological data to a fixed grid of a set size, covering a specified latitude and longitude area.

[0029] Further, in step S41, the sliding window extraction layer extracts a local meteorological field, including:

[0030] S411, set the window size; S412, dynamically adjust the window offset according to the predicted position of the typhoon center; S413, boundary constraint check: when the distance between any boundary of the window and the data global boundary is less than 10 grid units, automatically adjust the offset to ensure that the window is completely within the data boundary; S414, extract the local meteorological field: output a three-dimensional tensor, which is used as an MSF-Net deep learning model input unit, including spatial and temporal dimension meteorological field information.

[0031] In step S42, the multi-scale feature fusion network MSF-Net inference layer predicts the relative position offset of the typhoon center, including:

[0032] S421, based on the MSF-Net deep learning model input unit, the MSF-Net network backbone of the multi-scale feature fusion network MSF-Net inference layer performs an inference process;

[0033] S422, the multi-scale feature fusion network MSF-Net inference layer outputs the relative position index of the typhoon center;

[0034] In step S421, the MSF-Net network backbone performs an inference process, including:

[0035] S4211, primary feature encoding: extracting weather field texture features step by step through multi-level convolution operation;

[0036] S4212, deep feature enhancement: introducing Gaussian noise in the convolution layer; adding input features and output features through cross-layer connection;

[0037] S4213, global context fusion: integrating multi-scale features using convolution kernel and generating global feature vector;

[0038] S4214, feature compression: compressing global feature vector through fully connected network.

[0039] Further, in step S43, the path coordinate generation layer updates the longitude and latitude index and the longitude and latitude boundary index of the typhoon center, including:

[0040] S431, decoding relative position index: converting the position index in the relative window into longitude and latitude index;

[0041] S432, updating typhoon center coordinates: taking the decoded longitude and latitude index as the new typhoon center coordinates, and calculating the longitude and latitude boundary index of the typhoon center based on the longitude and latitude index of the typhoon center.

[0042] In a second aspect, based on the same aspect, the application also provides an artificial intelligence driven typhoon path dynamic tracking system, including a running environment, a data control program and a typhoon positioning program;

[0043] The running environment is a CPU running environment or a GPU running environment;

[0044] The data control program includes an initial coordinate processing layer, a meteorological data processing layer and a geographic coordinate mapping layer; the initial coordinate processing layer is used to determine the grid index and the boundary range according to the typhoon starting position, the meteorological data processing layer is used to fuse and grid resample multi-source meteorological data, and the geographic coordinate mapping layer is used to establish the bidirectional conversion relationship between geographic coordinates and grid index;

[0045] The typhoon positioning program includes a sliding window extraction layer, a multi-scale feature fusion network MSF-Net inference layer and a path coordinate generation layer; the sliding window extraction layer is used to extract local weather field, the multi-scale feature fusion network MSF-Net inference layer is used to predict the relative position offset of the typhoon center, and the path coordinate generation layer is used to convert the relative position into absolute coordinates and update the longitude and latitude index and the longitude and latitude boundary index.

[0046] Further, the multi-scale feature fusion network MSF-Net inference layer includes a primary feature encoding module, a deep feature enhancement module, a global context fusion module, a feature compression module and a relative position index prediction module;

[0047] The primary feature encoding module is used to extract the texture features of the meteorological field step by step, the deep feature enhancement module is used to introduce Gaussian noise in the convolution layer and add the input features and the output features through cross-layer connection, the global context fusion module is used to integrate the multi-scale features through the convolution kernel to generate a global feature vector, the feature compression module is used to compress the global feature vector through a fully connected network, and the relative position index prediction module is used to predict the relative position index.

[0048] Compared with the prior art, the artificial intelligence driven typhoon path dynamic tracking method and system has the following beneficial effects:

[0049] (1) The spatiotemporal sliding window tracking mechanism is adopted, the window position is dynamically adjusted, the typhoon core area is focused on, and the spatiotemporal local features are effectively captured.

[0050] (2) The convolutional neural network (CNN) is used to extract the meteorological field features, which can more accurately predict the mutation and turning of the typhoon path compared with the traditional method.

[0051] (3) The GPU is used to accelerate the training process, which significantly improves the calculation efficiency while ensuring the integrity of the features, and meets the real-time requirement.

[0052] (4) For the typhoon path tracking task, a special network structure including a Gaussian noise enhancement layer, a multi-level down-sampling module and a residual connection is designed, Gaussian noise is injected to improve the robustness of the model to data noise and enhance the generalization ability.

[0053] (5) Progressive feature compression gradually compresses the 11-channel input to 2-channel offset output through 5-level convolution operation, which reduces the calculation amount while retaining the feature information essential for typhoon path prediction.

[0054] (6) The geographic coordinate inverse mapping algorithm converts the grid index into latitude and longitude coordinates, supports non-uniform grid mapping, covers the area from 60°N to 0° in latitude and from 60°E to 180°E in longitude, and meets the demand of northwest Pacific typhoon tracking. BRIEF DESCRIPTION OF DRAWINGS

[0055] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application and are incorporated in and constitute a part of this application. The embodiments of the present application and their description are used to explain the application and do not constitute an improper limitation on the application. In the drawings:

[0056] Figure 1 The system overall flowchart described in the embodiments of the present application;

[0057] Figure 2 The system overall architecture diagram described in the embodiments of the present application;

[0058] Figure 3 A quality control system flowchart as described in embodiments of the application;

[0059] Figure 4 A path coordinate generation layer flowchart as described in embodiments of the application;

[0060] Figure 5 An MSF-Net inference layer flowchart as described in embodiments of the application;

[0061] Figure 6 A sliding window extraction layer flowchart as described in embodiments of the application;

[0062] Figure 7 A data post-processing layer flowchart as described in embodiments of the application;

[0063] Figure 8 A geographic coordinate mapping layer flowchart as described in embodiments of the application;

[0064] Figure 9 A meteorological data processing layer flowchart as described in embodiments of the application;

[0065] Figure 10 An initial coordinate processing layer flowchart as described in embodiments of the application;

[0066] Figure 11 A dynamic window tracking diagram as described in embodiments of the application.

[0067] Figure 12 A model structure diagram of the application. DETAILED DESCRIPTION

[0068] It should be noted that the embodiments in the application and the features in the embodiments can be combined with each other without conflict.

[0069] In the description of the present application, it needs to be understood that the terms "center", "longitudinal", "transverse", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, which is only for the convenience of describing the present application and simplifying the description, and does not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first", "second" and the like are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined with "first", "second" and the like can be explicitly or implicitly included one or more. In the description of the present application, unless otherwise specified, the meaning of "a plurality of" is two or more.

[0070] In the description of the present application, it needs to be understood that unless otherwise explicitly specified and limited, the terms "mounting", "connection", "connection" should be understood in a broad sense, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium; it can be the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood through specific circumstances.

[0071] The present application will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0072] As Figures 1 to 2 shown, an artificial intelligence driven typhoon path dynamic tracking system includes a running environment, a data control program and a typhoon positioning program; the running environment is a CPU running environment or a GPU running environment; the data control program includes an initial coordinate processing layer, a meteorological data processing layer and a geographic coordinate mapping layer; the typhoon positioning program includes a sliding window extraction layer, a multi-scale feature fusion network MSF-Net inference layer and a path coordinate generation layer. An artificial intelligence driven typhoon path dynamic tracking method includes the following steps:

[0073] S1, training a multi-scale feature fusion network MSF-Net model;

[0074] S2, based on the original numerical weather prediction data, the meteorological data processing layer performs meteorological data fusion and grid resampling, and outputs the processed numerical weather prediction data;

[0075] S3, based on the initial information of the typhoon, the initial coordinate processing layer determines the grid index and the boundary range of the typhoon center; the geographic coordinate mapping layer establishes the bidirectional conversion relationship between the geographic coordinate and the grid index;

[0076] S4, cyclically iteratively updating the typhoon path tracking coordinates;

[0077] S5, outputting the typhoon path tracking data, and performing Gaussian smoothing and quality control processing on the typhoon path tracking data.

[0078] The specific implementation is as follows:

[0079] I. Multi-scale feature fusion network MSF-Net model training

[0080] Data source: Adopt the typhoon path data (CMA best path set) in the northwest Pacific Ocean from 2020 to 2024, including the typhoon center longitude and latitude coordinates, and the corresponding meteorological field parameters at the moment.

[0081] Proportion:

[0082] Training set: 2020-2023 data (80%), used for model parameter learning;

[0083] Validation set: 2020-2023 data from January to August (20%), used for hyperparameter tuning;

[0084] Test set: 2024 data, used for final performance evaluation.

[0085] Data format: Each sample includes near-surface meteorological element field (such as sea level pressure, 10-meter UV wind field) and upper air meteorological element field (such as 700 hPa / 850 hPa UV wind field, potential height, relative vorticity, etc.) 11-channel spatiotemporal sequence data and corresponding typhoon center label. According to the above division, a total of 2920 data are finally obtained for training, verification and testing. After training and tuning, the model is saved as a pth format file.

[0086] II. System architecture

[0087] The system input includes two types of data sources: original numerical weather prediction data and typhoon initial information. The system adopts a double-program collaborative architecture, which processes data control program and typhoon positioning program in parallel through a cyclic iteration mechanism to realize dynamic tracking of typhoon path.

[0088] The data control program is mainly responsible for the input and output management of data, including three core processing layers: the initial coordinate processing layer determines the grid index and boundary range according to the typhoon coding position; the meteorological data processing layer fuses and resamples the multi-source meteorological data; the geographic coordinate mapping layer establishes the bidirectional conversion relationship between longitude and latitude and grid index. Through the cyclic updating mechanism, the program finally outputs the complete time series typhoon path tracking data.

[0089] The typhoon positioning program adopts a three-layer progressive processing architecture: the sliding window extraction layer extracts a 40x40 grid region centered on the current position of the typhoon; the multi-scale feature fusion network MSF-Net inference layer extracts features and predicts relative position offsets based on a deep learning model; and the path coordinate generation layer converts the relative positions into absolute coordinates and updates the boundary index. The program uses a closed-loop feedback mechanism, using the updated coordinates and boundary index obtained from the current prediction as input for the next time step, and iterates until all forecast horizons are predicted.

[0090] The data control program has the following levels of functions and interaction logic:

[0091] 2.1 Initial coordinate processing layer, as shown in Figure 10

[0092] Function: First, the initial latitude and longitude coordinates (lat_obs, lon_obs) of the typhoon are converted into initial latitude and longitude indices (lat_idx_init, lon_idx_init) on the 241x481 fixed grid using the relationship provided by the geographic coordinate mapping layer (see section 2.3). The initial latitude and longitude indices refer to the row and column numbers of the grid point in the global data array, both starting from 0.

[0093] Subsequently, a dynamic tracking area centered on the initial position of the typhoon is set. The boundary range of this area is defined by the initial latitude and longitude boundary indices, which are calculated as follows:

[0094] Latitude boundary index: lat_idx_min = lat_idx_init - 20, lat_idx_max = lat_idx_init + 20;

[0095] Longitude boundary index: lon_idx_min = lon_idx_init - 20, lon_idx_max = lon_idx_init + 20;

[0096] lat_idx_init: Initial latitude grid row number of the typhoon center. lon_idx_init: Initial longitude grid column number of the typhoon center. lat_idx_min: Minimum boundary row number of the tracking area in the latitude direction. lat_idx_max: Maximum boundary row number of the tracking area in the latitude direction. lon_idx_min: Minimum boundary column number of the tracking area in the longitude direction. lon_idx_max: Maximum boundary column number of the tracking area in the longitude direction.

[0097] ​The value "20" is set based on the fact that the sliding window size used in this system is 40×40 grid units (see section 2.5.1). The formula means that, with the initial typhoon location as the center, expanding outwards by 20 grid points in each direction creates a square area with sides of 40 grid units, which serves as the initial search range for subsequent sliding window extraction. This ensures that the initial window can completely cover the core vortex structure of the typhoon.

[0098] 2.2 Meteorological data processing layer, such as Figure 9 As shown:

[0099] Input: Receive raw numerical weather forecast data (6-hour interval numerical weather forecast model forecast data).

[0100] The processing includes:

[0101] 2.2.1 Meteorological data fusion:

[0102] Content: Integrates 11 key meteorological parameter channels, specifically covering near-surface meteorological elements (sea level pressure, 10-meter UV wind field) and upper-air meteorological elements (700hPa / 850hPa UV wind field, geopotential height, relative vorticity).

[0103] Resolution: The original data resolution is 0.25°×0.25° latitude and longitude grid.

[0104] 2.2.2 Grid resampling processing:

[0105] Method: Bilinear interpolation algorithm is used.

[0106] Target grid: Data is resampled to a fixed grid of 241×481, covering the latitude range of 60°N to 0° and the longitude range of 60°E to 180°E.

[0107] Objective: To ensure spatial consistency of meteorological data from different sources and resolutions, and to output processed numerical weather forecast data.

[0108] 2.3 Geographic coordinate mapping layer, such as Figure 8 As shown:

[0109] Method: Establish a bidirectional mapping relationship between latitude and longitude and grid index.

[0110] Implementation method: The conversion between latitude and longitude and grid index is completed quickly by pre-compiling a lookup table.

[0111] Performance: A single query takes less than 0.1ms, meeting real-time requirements.

[0112] 2.4 Data post-processing layer, such as Figure 7 As shown:

[0113] 2.4.1 Reverse geocoding

[0114] Function: Establish the bidirectional conversion relationship between geographic coordinates (latitude and longitude) and grid index.

[0115] The latitude and longitude range covered by the fixed grid is: latitude 60°N to 0°, longitude 60°E to 180°E. The grid resolution is 0.25°x0.25°, so there are (60° / 0.25°)=241 points in the latitude direction and (120° / 0.25°)=481 points in the longitude direction.

[0116] The formula for calculating the geographic coordinates (lat, lon) from the grid index (lat_idx, lon_idx) is as follows:

[0117] Latitude: lat = 60° - (lat_idx / 240) × 60°;

[0118] Longitude: lon = 60° + (lon_idx / 480) × 120°;

[0119] Explanation of the meaning of the formula parameters:

[0120] lat_idx and lon_idx: the index values of the grid points in the latitude and longitude directions, respectively.

[0121] Denominator 240 and 480: because there are 241 points from index 0 to index 240, but there are 240 intervals. The same applies to longitude. This calculation represents the equal division of the total number of intervals.

[0122] Latitude formula: the starting point is 60°N (when lat_idx=0), and as lat_idx increases, the latitude decreases linearly towards the equator (0°).

[0123] Longitude formula: the starting point is 60°E (when lon_idx=0), and as lon_idx increases, the longitude increases linearly towards 180°E.

[0124] Correspondingly, the formula for calculating the grid index (lat_idx, lon_idx) from the geographic coordinates (lat, lon) can be obtained by reversing the above formula, which is:

[0125] lat_idx = floor( (60° - lat) / 60° × 240 );

[0126] lon_idx = floor( (lon - 60°) / 120°× 480 ); where floor(·) is the floor function.

[0127] 2.4.2 Gaussian smoothing filter:

[0128] Parameter: σ = 1.5 grid units.

[0129] Objective: To smooth the predicted path, eliminate jitter that may occur in single-step prediction, and improve the consistency and reliability of the path.

[0130] 2.4.3 Quality control system, such as Figure 3 As shown:

[0131] Reasonableness check:

[0132] Coastline buffer zone: When the predicted location is more than 50km from the coastline, an early warning mechanism is triggered.

[0133] Abnormal movement speed: When the typhoon's movement speed exceeds 40m / s, it is considered abnormal data and will be marked or corrected.

[0134] 2.4.4 Output: Output the final typhoon path data.

[0135] The functions and interaction logic of each level of the typhoon location program are as follows:

[0136] 2.5 Sliding window extraction layer, such as Figure 6 As shown:

[0137] Input: Processed numerical weather forecast data, current typhoon center coordinates (initial latitude and longitude index or updated latitude and longitude index), and boundary index.

[0138] 2.5.1 Window Size Setting:

[0139] Size: 40×40 grid units.

[0140] Spatial range: Corresponding to an area of ​​approximately 1000km × 1000km, capable of covering the main impact area of ​​typhoons.

[0141] 2.5.2 Dynamic window offset adjustment:

[0142] Basis: The window offset is dynamically calculated based on the predicted positions of the typhoon centers from the previous two typhoons.

[0143] Objective: To ensure that the window is always focused on the core area of ​​the typhoon and to improve the targeting of feature extraction.

[0144] 2.5.3 Boundary constraint check:

[0145] Range Limitation: Ensure the grid indices of the window do not exceed the data range, i.e., 0 ≤ lat_idx < 241, 0 ≤ lon_idx < 481.

[0146] Treatment: When the window approaches the data boundary, automatically adjust the offset to avoid boundary crossing.

[0147] 2.5.4 Spatio-temporal Cube Construction:

[0148] Output: Form a 11×40×40 three-dimensional tensor as the MSF-Net deep learning model input unit.

[0149] Significance: The MSF-Net deep learning model input unit includes both spatial and temporal dimension information, which helps the model capture the dynamic evolution characteristics of typhoons.

[0150] 2.6 Multi-scale Feature Fusion Network MSF-Net Inference Layer, as shown in Figure 5

[0151] This layer receives the MSF-Net deep learning model input unit, performs inference through the MSF-Net network backbone, and outputs the relative position index.

[0152] 2.6.1 Primary Feature Encoding:

[0153] Structure: Extract texture features of meteorological fields step by step through 3 layers of convolution operations (Conv0-Conv2).

[0154] Parameter Variation: The number of output channels increases from 11 to 128 step by step, and the spatial resolution gradually decreases to 1 / 8 of the original size.

[0155] Activation Function: ReLU activation function is used to increase the non-linear expression ability of the network.

[0156] 2.6.2 Deep Feature Enhancement:

[0157] Gaussian Noise Injection: Introduce Gaussian noise (σ=0.1) in the 3rd and 4th convolution layers (Conv3-Conv4).

[0158] Residual Connection: Add input features and output features through cross-layer connection to prevent gradient disappearance and accelerate training convergence.

[0159] Purpose: Enhance the robustness of the model to input data noise and improve the stability of feature extraction.

[0160] 2.6.3 Global Context Fusion:

[0161] Method: The final convolution layer (Conv5) uses a 1×1 convolution kernel to integrate multi-scale features.​

[0162] Output: Generates a 1024-dimensional global feature vector to capture the overall features of the typhoon.

[0163] 2.6.4 Feature Compression:

[0164] Structure: The 1024-dimensional features are compressed through a 4-layer fully connected network (FC1-FC4).

[0165] Target dimension: Ultimately compressed to 128 dimensions, reducing computational load while retaining key information.

[0166] Activation function: Each layer is followed by a ReLU activation function to maintain the nonlinear characteristics of the network.

[0167] 2.6.5 Relative Position Index Prediction:

[0168] Output layer design: Instead of using an activation function, we directly regress the typhoon center within the relative window using MSF-Net deep learning, and output the relative position index (lat_idx_rel, lon_idx_rel).

[0169] Advantages: It avoids the nonlinear distortion that may be introduced by the activation function, thus improving the accuracy of location prediction.

[0170] 2.7 Path coordinate generation layer, such as Figure 4 As shown:

[0171] Input: relative position index, boundary index (initial latitude and longitude boundary index or updated latitude and longitude boundary index).

[0172] 2.7.1 Relative Position Index Decoding:

[0173] formula:

[0174] Latitude: lat_idx_pre = lat_idx_rel + lat_idx_min;

[0175] Longitude: lon_idx_pre = lon_idx_rel + lon_idx_min;

[0176] Where, lat_idx_rel: the relative latitude row number of the model-predicted typhoon center within the sliding window. lon_idx_rel: the relative longitude column number of the model-predicted typhoon center within the sliding window. lat_idx_min: the minimum latitude boundary row number of the current sliding window in the global grid. lon_idx_min: the minimum longitude boundary column number of the current sliding window in the global grid. lat_idx_pre: the predicted latitude row number of the typhoon center in the global grid. lon_idx_pre: the predicted longitude column number of the typhoon center in the global grid.

[0177] Objective: Convert the position index within the relative window to the updated latitude and longitude index.

[0178] 2.7.2 Typhoon center coordinate update:

[0179] Operation: Take the decoded updated latitude and longitude index lat_idx_pre and lon_idx_pre as the new typhoon center coordinates, and calculate the updated latitude and longitude boundary index accordingly.

[0180] Loop update: The updated coordinates and boundary index will be used as the input for the next loop update, realizing continuous tracking of the path.

[0181] The advantages and beneficial effects of the present application are as follows:

[0182] (1) Adopting a spatiotemporal sliding window tracking mechanism, focusing on the core area of the typhoon by dynamically adjusting the window position, effectively capturing the local spatiotemporal features.

[0183] (2) Using convolutional neural network (CNN) to extract meteorological field features, which can more accurately predict the mutation and turning of the typhoon path compared to traditional methods.

[0184] (3) Using GPU to accelerate the training process, significantly improving the computational efficiency while ensuring the integrity of the features, meeting the real-time requirements.

[0185] (4) For the task of typhoon path tracking, a special network structure including a Gaussian noise enhancement layer, a multi-level downsampling module, and a residual connection is designed, which injects Gaussian noise to improve the robustness of the model to data noise and enhances the generalization ability.

[0186] (5) Progressive feature compression gradually compresses the 11-channel input to a 2-channel offset output through 5-level convolution operations, reducing the computational load while retaining the essential feature information for typhoon path prediction.

[0187] (6) The geographic coordinate inverse mapping algorithm converts the grid index into latitude and longitude coordinates, supports non-uniform grid mapping, and covers the area from latitude 60°N to 0° and longitude 60°E to 180°E, meeting the needs of typhoon tracking in the Northwest Pacific.

[0188] Example 1

[0189] like Figure 12 As shown, a deep learning model architecture specifically designed for processing meteorological element data is presented. The processing flow begins with multi-channel meteorological element input data organized in a two-dimensional grid, with different colors representing the numerical distribution of different meteorological elements. The data first enters the deep feature encoding module, which uses carefully designed downsampling ratios of 11, 32, 64, 128, 256, 512, and 1024 to gradually reduce spatial resolution while extracting multi-level feature representations using operations such as convolutional and pooling layers, achieving hierarchical capture from local details to global features. Subsequently, the primary feature encoding module further refines and reduces the dimensionality of the deep features, laying the foundation for subsequent processing. The processing flow then enters the crucial global context fusion module, which fuses local features with overall meteorological model information through mechanisms such as global average pooling, effectively enhancing the model's understanding of the macroscopic structure of the atmospheric system. Finally, the feature compression module uses techniques such as multilayer perceptrons to compress high-dimensional features into compact latent representations and generates the typhoon center relative position index (lat_idx_rel, lon_idx_rel).

[0190] like Figure 11 As shown in the figure, the complete closed-loop process of the deep learning-based intelligent typhoon path tracking method in this invention is clearly illustrated. Its technical logic is as follows: First, the system takes the known initial state of the typhoon (center_init) as input, which includes precise latitude and longitude coordinates and the corresponding meteorological element observation area bounding box. Then, the multi-channel meteorological element field within the bounding box is input into a pre-trained multi-scale feature fusion network (MSF-Net). This network, through its unique encoder-decoder structure and feature pyramid design, can simultaneously capture the macroscopic environmental field features and local eye region structural features of the typhoon system, and outputs the predicted typhoon center position and range bounding box (center_next) for the next time period. This prediction result is then fed back to the model input as a new initial state, forming a recursive inference chain of "state input - feature extraction - position prediction - state update". Through this sequential iterative prediction mechanism, the system can automatically complete the continuous tracking of the typhoon's movement path until the end of the entire observation period, ultimately generating a continuous, smooth, and physically meaningful typhoon path trajectory.

[0191] The above merely provides the preferred embodiment of the present application, and is not used to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. An artificial intelligence-driven method for dynamic tracking of typhoon paths, characterized in that: Includes the following steps: S1. Train the MSF-Net model, a multi-scale feature fusion network; The MSF-Net model includes the following modules: a primary feature encoding module, used to extract texture features of the input meteorological field step by step through multi-level convolutional operations; a deep feature enhancement module, used to introduce Gaussian noise into the convolutional layers and add the input features and output features through cross-layer connections; a global context fusion module, used to integrate multi-scale features using convolutional kernels and generate a global feature vector; a feature compression module, used to compress the global feature vector through a fully connected network; and a relative position index prediction module, used to output the relative position index of the typhoon center. S2. Based on the original numerical weather forecast data, the meteorological data processing layer performs meteorological data fusion and grid resampling, and outputs the processed numerical weather forecast data. S3. Based on the initial typhoon information, the initial coordinate processing layer determines the grid index and boundary range of the typhoon center; the geographic coordinate mapping layer establishes a two-way conversion relationship between geographic coordinates and grid index. S4. Iterate and update the typhoon path tracking coordinates in a loop; S5. Output typhoon path tracking data and perform Gaussian smoothing and quality control processing on the typhoon path tracking data. In step S3, based on the initial typhoon information, the initial coordinate processing layer determines the grid index and boundary range of the typhoon center, including: Based on the typhoon's starting location, the initial coordinate processing layer generates the initial latitude and longitude index of the typhoon center and the initial latitude and longitude boundary index. In step S4, the typhoon path tracking coordinates are updated iteratively, including: S41. Based on the processed numerical weather forecast data, the latitude and longitude index of the typhoon center, and the latitude and longitude boundary index, a sliding window extraction layer is used to extract the local meteorological field. S42. Based on the extracted local meteorological field, the MSF-Net inference layer of the multi-scale feature fusion network predicts the relative position shift of the typhoon center. S43. Based on the predicted relative position of the typhoon center, the path coordinate generation layer updates the latitude and longitude index and latitude and longitude boundary index of the typhoon center. S44. The updated latitude and longitude indexes of the typhoon center and the latitude and longitude boundary indexes are fed back into the sliding window extraction layer for the next round of iteration.

2. The artificial intelligence-driven dynamic tracking method for typhoon paths according to claim 1, characterized in that: In step S1, the multi-scale feature fusion network MSF-Net model is trained, including: S11. Divide the typhoon path data into training set, validation set and test set; S12. Using the training set, validation set, and test set, the MSF-Net model of the multi-scale feature fusion network is learned, optimized, and evaluated. S13. Save the MSF-Net model of the multi-scale feature fusion network.

3. The method for dynamic tracking of typhoon paths driven by artificial intelligence according to claim 1, characterized in that: In step S2, the meteorological data processing layer performs meteorological data fusion and grid resampling, including: S21. Meteorological data fusion: Integrating meteorological parameter channels covering near-surface meteorological element fields and upper-air meteorological element fields; S22. Grid resampling processing: The meteorological data is resampled to a fixed grid of a set size using a bilinear interpolation algorithm, covering the specified latitude and longitude area.

4. The artificial intelligence-driven dynamic tracking method for typhoon paths according to claim 1, characterized in that: In step S41, the sliding window extraction layer extracts the local meteorological field, including: S411, Set window size; S412. Dynamically adjust the window offset based on the predicted location of the typhoon center; S413 Boundary constraint check: When the distance between any boundary of the window and the boundary of the global data area is less than 10 grid units, the offset is automatically adjusted; S414. Extract local meteorological field: Output a three-dimensional tensor and use it as the input unit of the MSF-Net deep learning model. The input unit of the MSF-Net deep learning model includes meteorological field information in spatial and temporal dimensions.

5. The artificial intelligence-driven dynamic tracking method for typhoon paths according to claim 1, characterized in that: In step S42, the MSF-Net inference layer of the multi-scale feature fusion network predicts the relative position shift of the typhoon center, including: S421. Based on the input unit of the MSF-Net deep learning model, the MSF-Net network backbone of the multi-scale feature fusion network MSF-Net inference layer performs the inference process; S422, The MSF-Net inference layer of the multi-scale feature fusion network outputs the relative position index of the typhoon center; In step S421, the MSF-Net network backbone performs the inference process, including: S4211, Primary Feature Encoding: Extracting meteorological field texture features step by step through multi-level convolution operations; S4212, Deep Feature Enhancement: Introduce Gaussian noise into the convolutional layer; add the input features and output features through cross-layer connections; S4213, Global Context Fusion: Convolutional kernels are used to integrate multi-scale features and generate a global feature vector; S4214. Feature Compression: Compress the global feature vector through a fully connected network.

6. The artificial intelligence-driven dynamic tracking method for typhoon paths according to claim 1, characterized in that: In step S43, the path coordinate generation layer updates the latitude and longitude indexes of the typhoon center and the latitude and longitude boundary indexes, including: S431. Decode relative position index: Convert the position index within the relative window into a latitude and longitude index; S432. Update typhoon center coordinates: Use the decoded latitude and longitude index as the new typhoon center coordinates, and calculate the latitude and longitude boundary index of the typhoon center based on the latitude and longitude index of the typhoon center.

7. An AI-driven typhoon path dynamic tracking system, applied to the AI-driven typhoon path dynamic tracking method described in any one of claims 1-6, characterized in that: This includes the operating environment, data control program, and typhoon location program; The operating environment is either a CPU operating environment or a GPU operating environment; The data control program includes an initial coordinate processing layer, a meteorological data processing layer, and a geographic coordinate mapping layer. The initial coordinate processing layer is used to determine the grid index and boundary range based on the typhoon's starting location. The meteorological data processing layer is used to fuse multi-source meteorological data and perform grid resampling. The geographic coordinate mapping layer is used to establish a bidirectional conversion relationship between geographic coordinates and grid indexes. The typhoon location procedure includes a sliding window extraction layer, a multi-scale feature fusion network MSF-Net inference layer, and a path coordinate generation layer. The sliding window extraction layer is used to extract the local meteorological field, the multi-scale feature fusion network MSF-Net inference layer is used to predict the relative position shift of the typhoon center, and the path coordinate generation layer is used to convert the relative position into absolute coordinates and update the latitude and longitude index and latitude and longitude boundary index.

Images