A High-Precision Intelligent Prediction Method for Typhoon Tracks Based on Multi-Source Asymmetric Fusion Networks

Abstract

This invention discloses a high-precision intelligent prediction method for typhoon paths based on a multi-source asymmetric fusion network, belonging to the field of intelligent meteorological disaster prediction and multi-source data fusion technology. The method includes: acquiring historical typhoon path data and reanalyzing meteorological data; performing rolling pruning based on the prediction anchor point of the previous moment to produce standardized sample pairs for the current moment; performing asymmetric multi-source encoding on these pairs to generate 2D and 3D branch features, and performing mutation perception and geometric consistency interaction on them to generate a basic prediction of the typhoon path; monitoring specific meteorological events, and performing event-conditional micro-perturbation amplification and path coupling based on the basic prediction of the typhoon path to generate a final prediction of the typhoon path; using the final prediction of the typhoon path as the prediction anchor point for the next moment and incorporating it into the inference trajectory sequence. This invention can improve the model's ability to capture complex meteorological changes.

Description

Technical Field

[0001] This invention belongs to the field of intelligent prediction of meteorological disasters and multi-source data fusion technology, and in particular, it is a high-precision intelligent prediction method for typhoon paths based on a multi-source asymmetric fusion network. Background Technology

[0002] Typhoons, as one of the most severe natural disasters affecting the Northwest Pacific and coastal areas of China, exhibit highly nonlinear and uncertain paths, influenced by a variety of complex factors such as atmospheric circulation, topography, and sea surface temperature. With the development of meteorological observation and remote sensing technologies, multi-source heterogeneous data (such as historical typhoon paths, meteorological reanalysis data, and satellite imagery) have provided a rich information foundation for typhoon path prediction.

[0003] Existing technologies for typhoon track prediction include statistical models, numerical weather prediction models, and statistical-dynamic hybrid models. Statistical models, such as the CLIPER method, can achieve rapid predictions based on the regression relationship between historical tracks and meteorological factors, but they have poor adaptability to extreme tracks and abrupt events. Numerical weather prediction models are based on atmospheric dynamic equations and have high theoretical accuracy, but they are sensitive to initial fields and parameters, consume significant computational resources, and still suffer from large track biases and insufficient real-time performance under complex climate conditions. Furthermore, in recent years, deep learning techniques have been employed for typhoon track prediction.

[0004] However, existing research still faces challenges such as insensitivity of predictive models to path abrupt changes, insufficient ability to capture perturbations from specific physical events (e.g., ERC), and poor generalization ability due to data leakage risks. Existing models struggle to efficiently integrate multi-scale information simultaneously. Therefore, further research and innovation are needed to address these issues in current technologies. Summary of the Invention

[0005] Purpose of the invention: In view of the above-mentioned problems of the prior art, this application provides a high-precision intelligent prediction method for typhoon paths based on multi-source asymmetric fusion networks.

[0006] Technical solution: According to the first aspect of this application, a high-precision intelligent prediction method for typhoon paths based on multi-source asymmetric fusion networks is provided, comprising:

[0007] Acquire historical typhoon path data and reanalysis meteorological data, perform rolling pruning based on the previous moment's forecast anchor point, and produce standardized sample pairs for the current moment;

[0008] Asymmetric multi-source coding is performed on standardized sample pairs to generate 2D and 3D branch features, and mutation-aware and geometrically consistent interactions are performed on them to generate basic predictions of typhoon paths.

[0009] By monitoring specific meteorological events and performing event-conditional micro-perturbation amplification and path coupling based on the basic prediction of typhoon paths, the final prediction of typhoon paths is generated.

[0010] The final prediction of the typhoon's path is used as the anchor point for the next moment's prediction and incorporated into the inference trajectory sequence.

[0011] According to a first aspect of this application, a high-precision intelligent prediction system for typhoon paths based on a multi-source asymmetric fusion network is provided, comprising:

[0012] processor;

[0013] Memory, which stores computer-executable instructions;

[0014] When the instruction is executed by the processor, it causes the system to perform the method as described in any of the first aspects of this application.

[0015] Beneficial effects: The present invention solves the problems of prediction models being insensitive to path mutations, having insufficient ability to capture disturbances of specific physical events (such as ERC), and having poor generalization ability due to the risk of data leakage. The relevant technical effects will be described in detail below with reference to specific embodiments. Attached Figure Description

[0016] Figure 1 A flowchart of a high-precision intelligent prediction method for typhoon paths based on a multi-source asymmetric fusion network provided in this application embodiment.

[0017] Figure 2 A flowchart for monitoring specific meteorological events provided in this application embodiment.

[0018] Figure 3 This is a flowchart illustrating an example of performing mutation-aware and geometrically consistent interaction on 2D branch features and 3D branch features, provided as an embodiment of this application.

[0019] Figure 4 This is another flowchart illustrating mutation-aware and geometrically consistent interaction between 2D branch features and 3D branch features, provided as an embodiment of this application.

[0020] Figure 5 A flowchart illustrating the rate of change of heading rate based on spherical geometric symmetry difference calculation provided in this application embodiment. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0022] It should be noted that the terms "first," "second," etc., in the specification and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "including" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0023] In some embodiments, a typhoon track prediction method based on Symmetric Fusion Network (MSAFN) is provided for comparison. One technical solution (such as the MSAFN model) employs a dual-branch network backbone. Specifically, this solution uses one branch (e.g., a bidirectional long short-term memory network Bi-LSTM) to process two-dimensional (2D) time-series features such as historical typhoon trajectories and intensities, and uses another branch (e.g., a three-dimensional convolutional neural network 3D-CNN) to process three-dimensional (3D) meteorological field data.

[0024] In the feature fusion stage, this scheme employs a multi-layered interaction mechanism. For example, a low-order feature interaction module (LFIM) is used to linearly project and stitch 2D and 3D features to capture linear correlations; a high-order nonlinear interaction module (HNIM) (e.g., using bilinear pooling) is used to achieve deep coupling of features. Furthermore, this scheme may also employ a general spatiotemporal attention mechanism to dynamically allocate weights across spatiotemporal regions.

[0025] However, the inventors recognize that the above technical solution has at least the following drawbacks:

[0026] Specifically, it lacks the ability to sense sudden changes; the LFIM and HNIM fusion mechanism it uses is a general, data-driven fusion method, lacking a specific sensing mechanism for sudden changes in typhoon paths (such as a sudden northward or westward turn). This results in larger prediction errors when the typhoon path undergoes drastic and nonlinear changes.

[0027] Furthermore, there is a lack of event-specific correction; the scheme does not explicitly consider the small but critical perturbation effects of specific meteorological events (such as eyewall replacement cycles, ERC) on the path. Path sway caused by events such as ERC are small in scale and easily smoothed out in coarse-grid 3D meteorological data, causing the model to fail to capture this type of physical process.

[0028] Based on this, there is a risk of data leakage; the proposed solution does not clearly define data pruning strategies during training and inference. If the model uses the future real typhoon location as the anchor point for pruning and reanalyzing the meteorological field during training or inference, it will cause data leakage, resulting in seemingly high model evaluation results (such as average distance error), but the actual performance will decrease in actual deployment (where the future real location cannot be known).

[0029] To address the aforementioned shortcomings in existing technologies, such as the predictive models' insensitivity to path abrupt changes, insufficient ability to capture disturbances from specific physical events (e.g., ERC), and poor model generalization ability due to data leakage risks, this paper combines... Figures 1 to 5 The present invention will be specifically described through the following embodiments.

[0030] Example 1 provides a general exemplary scheme for a high-precision intelligent prediction method for typhoon paths based on multi-source asymmetric fusion networks. This method keeps training and inference consistent and introduces a dedicated module to handle path abrupt changes and specific event disturbances.

[0031] Specifically, the method of this embodiment includes the following steps:

[0032] Step S101: Obtain historical typhoon path data and reanalysis meteorological data, perform rolling cropping based on the previous moment's prediction anchor point, and produce standardized sample pairs for the current moment.

[0033] Historical typhoon track data can include scalar time series data such as the typhoon center's location (latitude and longitude), central pressure, and maximum wind speed at past moments. Reanalysis meteorological data refers to three-dimensional meteorological field grid data, such as that provided by ERA5 or NCEP / NCAR, containing multiple isobaric layers (e.g., 1000hPa, 850hPa, 500hPa, 250hPa) and multiple variables (e.g., temperature T, geopotential height Z, relative humidity RH, zonal wind U, meridional wind V). The prediction anchor point refers to the typhoon center location output by the model when performing a prediction task at the previous time point (e.g., t-6h). This contrasts with the true anchor point, which is the actual location of the typhoon observed by the meteorological department at time (t). Rolling clipping refers to the system extracting meteorological tensors for a local area (e.g., a 5°×5° area centered on the prediction anchor point) from the global reanalysis meteorological data during the inference (prediction) phase, using the prediction anchor point (not the true anchor point) as the geographic center.

[0034] In real-world operational forecasting scenarios, the actual anchor points of typhoons at the current time (t) and future times are unknown. Therefore, the model must rely on its own forecast output to continuously construct subsequent input data. Strictly using the forecast anchor point from the previous time step during the inference phase ensures that the model's forecasting process is self-consistent, closed-loop, and consistent with real-world application scenarios.

[0035] As an alternative, during the training phase, an anti-leakage training scheduling strategy can be adopted. This involves switching between the real anchor point and the predicted anchor point from the previous time step with a certain probability (e.g., dynamically decreasing from 0.9 to 0.5 with each training epoch). This accelerates model convergence while improving its robustness to its own prediction errors. Standardized sample pairs refer to the training or inference data used as input to the model after preprocessing the cropped meteorological tensor and historical path data, such as Z-score normalization, outlier handling, and missing data repair.

[0036] Step S102: Perform asymmetric multi-source encoding on the standardized sample pairs to generate 2D branch features (F... _2D ) and 3D branching features (F _3D And perform mutation-aware and geometrically consistent interaction on its (2D branch features and 3D branch features) to generate a basic prediction of the typhoon track (p # _24h ),in, # The symbol is _hat.

[0037] Asymmetric multi-source coding refers to using network branches with different structures and computational complexities to process heterogeneous multi-source data separately. For example, for 2D path data (K...) with a small data volume but strong temporal sequence... _t ), using a relatively lightweight encoder (such as a two-layer Bi-LSTM) to generate 2D branch features; for 3D reanalysis meteorological tensors (G) with large amounts of data (such as T×L×H×W×C), _t ), using computationally intensive encoders (such as multi-scale 3D convolutional networks) to generate 3D branch features.

[0038] Among them, the Mutation Sensing and Geometrically Consistent Interaction (MAGI) replaces the general fusion module. Standard data-driven fusion (such as HNIM) struggles to capture low-probability but high-impact path abrupt changes (such as sudden typhoon changes). Therefore, the MAGI module (described in subsequent embodiments) includes a mutation precursor identifier to assess mutation probabilities within local time windows, dynamically enhancing the response to mutation signals using mutation gates. Furthermore, a Geometrically Consistent Interaction (GCA-LRB) approach is employed to perform physically intuitive fusion of 2D and 3D features on a spherical tangent plane, generating a fundamental typhoon path prediction highly sensitive to atmospheric circulation and the typhoon's own kinematics.

[0039] Step S103: Monitor specific meteorological events, perform micro-perturbation amplification and path coupling based on the basic prediction of the typhoon path, and generate the final prediction of the typhoon path (p). # _final ) Among them, specific meteorological events refer to physical processes that are known to affect the typhoon track, but whose scale may be smaller than the resolution of the 3D meteorological field, such as the eyewall replacement cycle (ERC).

[0040] In this embodiment, the system includes parallel event monitors for continuous analysis of input data (G). _t K _t The system detects whether a precursor to ERC (e.g., a double-loop structure or Rmax bimodality) has appeared. The system activates the event-conditionalized micro-perturbation amplification and path coupling (MPAN-Geo) module (described in subsequent embodiments) only when an ERC event is detected. Micro-perturbation amplification refers to this module cropping a tiny (e.g., 10km × 10km) central micro-patch around the current location and applying super-resolution reconstruction to amplify the perturbation signal caused by ERC at a small scale, extracting ERC micro-perturbation features. Path coupling refers to inputting these micro-perturbation features into a dedicated orientation-preserving mapping module (GCA-LRB-ERC) to generate a path deflection increment (Δp) representing the contribution of the ERC perturbation. # _ERC Based on this, the deflection increment is coupled with the baseline prediction through vector synthesis to obtain the final prediction of the typhoon's path. The baseline prediction mainly reflects the path trend determined by large-scale circulation (such as the subtropical high) and mesoscale systems (such as the typhoon's main circulation). Based on this, the path oscillation caused by specific small-scale physical events (such as ERC) is corrected, improving the physical fidelity and final accuracy of the prediction. If no specific event is detected, the final prediction equals the baseline prediction.

[0041] Step S104: Use the final prediction of the typhoon path as the prediction anchor point for the next moment and incorporate it into the inference trajectory sequence.

[0042] At time t, the model outputs a final prediction. At the next inference time t+1, this final prediction will be used as the prediction anchor for the previous time, and will be used to prune the reanalysis meteorological data at time t+1. Integrating the inference trajectory sequence refers to collecting and connecting the final predictions output at each time (t, t+1, t+2, ..., t+n) to form a complete prediction path, for example, for the next 24 or 48 hours.

[0043] It should be understood that this embodiment constitutes a closed-loop inference strategy to prevent data leakage, ensuring consistency in the anchor point pruning mechanism during the training phase (anti-leakage scheduling) and the inference phase, thereby guaranteeing the operation of the model in actual business.

[0044] Example 2 provides an exemplary scheme for asymmetric multi-source coding and low-order alignment for efficient and differentiated processing of heterogeneous 2D (time series) and 3D (temporal and atmospheric field) data.

[0045] In this embodiment, the asymmetric multi-source coding step includes:

[0046] Step S201: Apply multi-scale 3D convolution to the 3D reanalysis meteorological tensor in the standardized sample pair, perform causal time compression, compress its (3D reanalysis meteorological tensor) time dimension into a shorter summary dimension, and generate 3D branch features.

[0047] Specifically, the input 3D reanalysis meteorological tensor (e.g., the original 3D reanalysis meteorological tensor) has a shape of T2×L×H×W×C. Here, T2 is the historical time window step size (e.g., T2=12 steps, corresponding to 72 hours); L is the number of isobaric layers (e.g., L=4, corresponding to 1000, 850, 500, 250 hPa); H and W are the horizontal grid resolutions (e.g., corresponding to a 5°×5° geographic area); and C is the number of meteorological variable channels (e.g., C=5, corresponding to temperature, geopotential height, relative humidity, zonal wind, and meridional wind). Multi-scale 3D convolution is used, optionally employing a set of convolutional kernels with different receptive fields to process the input tensor in parallel. For example, this kernel set may include {3×3×3, 5×5×3, 7×7×3} to simultaneously capture meteorological patterns at different spatial scales. Optionally, this 3D convolutional encoder employs residual connections and group normalization to construct a deeper network structure and improve gradient propagation.

[0048] Causal time compression refers to reducing the temporal resolution of the 3D branch while maintaining temporal causality (i.e., the output at time t depends only on information up to and including time t). Asymmetric processing is employed because the evolution of 3D meteorological fields (such as large-scale circulation) is typically slower than the instantaneous changes in 2D paths; time compression reduces computational redundancy.

[0049] Optionally, this compression can be achieved through a 1×1×k causal dilation convolution along the time dimension. For example, different dilation rates such as {1, 2, 3} can be set to compress a time series of T2=12 steps into a segmented summary of T3=4 steps, generating 3D branch features.

[0050] Step S202, and the asymmetric multi-source coding further produces a low-order aligned representation; a first-order linear projection is applied to the summaries of the 2D and 3D branch features to generate the representation, preserving the common scale across modalities and suppressing noise dimensions. Alternatively, a first-order linear projection is applied to the summaries of the 2D and 3D branch features to preserve the common scale across modalities and suppress noise dimensions, producing a low-order aligned representation. The 2D branch features are generated from the 2D input sequence (including scalar time series such as historical trajectories and intensity).

[0051] As an optional implementation, to handle the periodicity of angles (such as heading) (e.g., the proximity of 0° and 360°), a sin / cos dual-channel expansion can be optionally performed on the heading before feeding it into the 2D branch encoder (e.g., Bi-LSTM) to avoid numerical jumps caused by angle discontinuities. Summarizing 2D branch features refers to reducing the temporal 2D branch features to a vector (e.g., obtaining a 2D summary vector by concatenating time-averaged features with the last time step). Summarizing 3D branch features refers to reducing the 3D branch features (shape T3×h3×H×W) to a vector (e.g., obtaining a 3D summary vector through global average pooling). Since simply concatenating the 2D and 3D summary vectors and feeding them into a non-linear activation function (e.g., ReLU) can result in a high-dimensional (e.g., h3=512) 3D summary vector and its contained noise dimension potentially overwhelming the low-dimensional (e.g., h2=256) 2D summary vector signal, this can lead to a high-dimensional (e.g., h3=512) 3D summary vector. Therefore, in this embodiment, two independent linear projection matrices W2 and W3 are optionally used to map the 2D summary vector and the 3D summary vector to a common low-dimensional feature space (e.g., d=256).

[0052] Furthermore, after projection, stitching, and / or layer normalization, no nonlinear activation function is applied. This linear interaction forces the model to learn the common, linear scale information between the two modes (trajectory kinematics and atmospheric environment field), suppressing the high-frequency noise dimension specific to each mode. The resulting low-order aligned representation can be used for subsequent regression prediction or as an auxiliary loss term to improve the model's stability and generalization ability.

[0053] Example 3 describes an optional implementation of Mutation Awareness and Geometric Consistent Interaction (MAGI) to generate basic predictions of typhoon paths, addressing the problem of model insensitivity to path mutations.

[0054] Step S301: Based on the standardized sample pairs at the current moment, calculate the kinematic features and environmental prior features, and combine them into a mutation feature vector. The kinematic features include the rate of change of heading calculated based on spherical geometric symmetry difference, and the environmental prior features include vertical wind shear and geopotential height time difference. Specifically, when calculating the rate of change of heading and the rate of change of heading, an angle-around normalization function is applied to the time difference of heading and the time difference of heading rate of change. The angle-around normalization function normalizes the time difference to a preset angle range, avoiding numerical jumps caused by azimuth rotation. The normalized time difference is then used to calculate the rate of change of heading.

[0055] One of the inputs to this step is the kinematic feature sequence K_t, K from the standardized sample pairs. _t It contains scalar information for all moments within a historical time window (e.g., T2=12 steps), such as historical location, central pressure, maximum wind speed, and speed and heading calculated from the historical location. The mutation feature vector produced in this step is based on K... _t and G _t Specific information in the (3D reanalysis meteorological tensor) (such as recent heading, wind shear, and geopotential height) is used through symmetric differencing and higher-order derivatives (such as the rate of change of heading, jerk). _t The low-dimensional feature vectors are derived and aggregated through methods such as physical priors (e.g., wind shear) and are specifically used to characterize mutations.

[0056] Correspondingly, kinematic characteristics refer to the geometric features of the typhoon's historical motion. To improve sensitivity to sudden changes in path, this embodiment optionally calculates higher-order derivative features. Specifically, symmetric differencing refers to using data from times t-Δ and t+Δ (e.g., Δ = 6 hours) when calculating features at time t. This method is less sensitive to noise and more stable than backward differencing (t and t-Δ). When calculating kinematic characteristics, especially when calculating the time difference of the heading and the time difference of the rate of change of heading, an angle wrap normalization function (e.g., wrap_π) is applied. This function normalizes the input time difference value (e.g., the difference is 340° when the typhoon changes from 350° to 10°) to a preset angle range (e.g., (-π, π], i.e., -180° to +180°, which would normalize 340° to -20°). This avoids numerical jumps caused by azimuth wrapping, ensuring that the calculated rate of change of heading (-1.67° / h) and its rate of change (jerk) are accurate. _t The value is a physically real, not an erroneous maximum (28.3° / h). (See subsequent examples for specific numerical calculations). Environmental a priori characteristics refer to key atmospheric physical quantities selected based on meteorological a priori knowledge and highly correlated with abrupt changes in typhoon path (such as the North Pole).

[0057] Specifically, the calculation of environmental prior features includes: calculating the vector wind difference modulus between the 850hPa and 250hPa isobars from the reanalysis meteorological data contained in the standardized sample pairs to obtain the vertical wind shear; and calculating the symmetric time difference of the geopotential height of the 500hPa isobar to obtain the geopotential height time difference (dZ_dt).

[0058] As an alternative, to enhance the perception of the typhoon's dynamic structure, prior environmental features can further include the circumferential mean and variance of the relative vorticity ζ850 calculated from the 850 hPa isobaric layer. Based on this, the kinematic features obtained above are combined with the prior environmental features (e.g., spliced) to form abrupt change feature vector f. _t .

[0059] Step S302: Input the 2D branch features and mutation feature vectors into the mutation flow encoder.

[0060] The mutation flow encoder can be built on a recurrent neural network unit (RNNCell), such as a long short-term memory network (LSTM) unit.

[0061] Because conventional LSTM units (or GRU units) use the same update intensity for candidate states at all time steps, and typhoon paths are smooth most of the time, with abrupt changes being sparse and transient, the mutation flow encoder uses a different update intensity at the input gate (i) of the recurrent neural network unit. _gate In addition to ), a mutation gate (g) is introduced. _mut Mutation gates are determined by mutation feature vectors (f). _t ) control (e.g., g) _mut =sigmoid(W _m •f _t +b _m ), where W _m and b _m (These are learnable parameters) used to dynamically amplify candidate states (c) within a cell. t ~ The update intensity of ) is represented by the mutation flow. Optionally, this dynamic amplification mechanism can be implemented by the following formula:

[0062] ;

[0063] Among them, c _t It is the cell state at the current time t; c _t-1 It is the cell state at the previous moment; f _gate It is the Gate of Oblivion; _gate It is an input gate; c t ~ It represents a candidate state; ⊙ indicates element-wise multiplication. It can be seen that when the mutation eigenvector indicates that a mutation is about to occur, g..._mut The model approaches 1 for candidate state c. t ~ The update intensity is amplified; while during the smoothing period, g _mut When the value approaches 0, the model inhibits updates and relies more on historical memory c. _t-1 .

[0064] Based on this, the encoder produces a mutation flow representation that is highly sensitive to abrupt signals (also known as a mutation-aware 2D representation F). _2D_m In some alternative implementations, the encoder can be implemented as a two-stream asynchronous structure. That is, the system maintains a mutation stream (e.g., MS-LSTM) and a regular smoothing stream (e.g., standard Bi-LSTM) in parallel. The final mutation-aware 2D representation F _2D_m Based on the outputs of the two streams above, the mutation probability P is estimated based on the mutation feature vector. _mutation (t) Perform dynamic weighted fusion (w) _mut •F _2D_m +w _smooth •F _2D_s It is formed by ) . Among them, w _mut with w _smooth For stream fusion weights, F _2D_s This is a smooth flow representation.

[0065] Step S303: For 3D branch features (F... _3D An isobaric layer selection gate is applied to allow the Top-K key layers to pass, generating a sparse 3D representation. This step (also known as PL-Gate) is an attention mechanism applied to the 3D branch features. Since not all isobaric layers contribute equally to path prediction...

[0066] Optionally, the gate calculates an importance score z for each isobaric level (e.g., 1000 hPa, 850 hPa, etc.). _level The score can be based on a 3D feature summary of that layer (e.g., pool(F)). _3D_level ) and mutation feature vector f _t Joint decision (e.g., z) _level =q T •[pool(F) _3D_level ), f _t ]; where q T It is the transpose of the learnable query vector q. The dimension of this vector q is the same as the concatenated features [pool(F _3D_level The dimensions of f_t are matched. Through backpropagation training, the vector q learns how to query and evaluate which combination of isobaric layer and which mutation feature is most important for prediction.

[0067] Next, the scores z of all layers are obtained by using a softmax function with temperature τ (τ∈[0.5,2]). _level Convert to probability weight α _level .

[0068] One method for generating sparse 3D representations is hard selection, which means only retaining (allowing) the weight α. _level The top-K layers (e.g., K=1 or 2); or soft selection, i.e., using α. _level As weights, the features of all isobaric layers are weighted and summed (expectation pooling).

[0069] Step S304: Based on the position sequence in the standardized sample pairs, construct a geometric orientation code on the spherical tangent plane. The geometric orientation code includes a unit tangent vector and a tangential velocity vector. This establishes a local coordinate system that conforms to physical intuition for subsequent physical fusion.

[0070] Specifically, at position p of the typhoon at the current moment t _t Define a local spherical tangent plane. Based on historical position sequences, calculate the current movement trend of the typhoon using spherical geometric differentiation. Unit tangent vector τ. _t It is a unit vector, and its direction defines the instantaneous tangential direction (i.e., heading) of the typhoon's path at that point. Tangential velocity vector v _t It is the instantaneous velocity vector of the typhoon on that tangential plane (its direction is related to τ). _t Consistent, amplitude is instantaneous velocity). Both (τ) _t v _t Together they form a geometric direction code d _t .

[0071] Step S305: The catastrophe flow representation, sparse 3D representation, and geometric direction encoding are fed into a direction-preserving low-rank bilinear interaction module (GCA-LRB). This module fuses 2D and 3D features and can replace the general fusion module. The contributions of 2D features (catastrophe flow representation, representing the typhoon's own kinematics and inertia) and 3D features (sparse 3D representation, representing the steering airflow imposed by the large-scale environmental field) to the final path are physically different and should not be simply concatenated.

[0072] Optionally, GCA-LRB implements a direction-preserving fusion method. The interaction module projects the abrupt flow representation onto the unit tangential vector direction (i.e., the contribution of the abrupt flow representation to the inertial direction) and the sparse 3D representation onto the tangential velocity vector direction (i.e., the contribution of the sparse 3D representation to the propulsion direction).

[0073] Optionally, this interaction can be implemented in a low-rank bilinear form. Let x = catastrophe flow representation and y = sparse 3D representation, then the interaction B(x, y) can be represented as:

[0074] B(x, y) = Σ _r=1..R (U) _r •x,τ _t • (V) _r •y,v _t );

[0075] Where Σ represents the summation of rank r from 1 to R (e.g., R∈[4,8]); U _r and V _r It is a learnable projection matrix (basis vector); (a, b) represents the inner product (i.e., projection) of vectors a and b. R is the rank and R∈[4,8], U _r and V _r τ is the projection matrix shared across the isobaric layer dimensions. _t Let v be the unit tangent vector representing the path inertial direction. _t The tangential velocity vector represents the propulsion direction, achieving a physically consistent fusion of 2D kinematic features and 3D environmental field features under spherical geometric constraints.

[0076] In one alternative implementation, to reduce the number of parameters, the projection matrix U... _r and V _r The parameters are shared across all isobaric layer dimensions. The interaction result B(x, y) can be further subjected to, for example, signed-sqrt and L2 normalization, to produce the final geometrically consistent representation.

[0077] Step S306: Generate a basic prediction of the typhoon path based on the geometrically consistent representation.

[0078] Specifically, the geometrically consistent representation (optionally fused with low-order aligned representation) is fed into a regression head (e.g., a multilayer perceptron MLP) to output the predicted longitude and latitude for the future (e.g., 24 hours). This output serves as the base prediction for the typhoon's path. Furthermore, to support subsequent training strategies, the module in this embodiment can also additionally output: the mutation probability P estimated based on the mutation feature vector. _mutation (t) and focal-like weights w used for imbalanced sample learning _focal A parallel uncertainty estimation head is used to output the prediction uncertainty σ of the basic forecast. _24h .

[0079] Example 4 describes an exemplary scheme for event-conditional ERC micro-perturbation amplification and path coupling (MPAN-Geo) to correct typhoon paths. Since the basic forecast of typhoon paths is mainly determined by large and medium-scale circulation, certain specific small-scale physical events (such as eyewall replacement) can cause small-scale but critical perturbations. This example is used to explicitly capture such perturbations.

[0080] Step S401: Using standardized samples at the current moment, monitor eyewall replacement precursors and generate ERC trigger flags. Among them, eyewall replacement precursors (ERC) are dynamic processes within the typhoon, usually manifested as the formation of the outer eyewall and its gradual replacement of the inner eyewall. This process is often accompanied by short-term oscillations in the path.

[0081] In this embodiment, the monitoring step can optionally be implemented by analyzing and reanalyzing meteorological tensors and kinematic characteristics. As an optional implementation, the monitor can calculate the radius-tangential wind mean profile I(R) of the typhoon core area (e.g., within a radius of 10 km to 100 km).

[0082] On the one hand, the probability P of the double-ring structure is calculated by searching for two local maxima on the profile I(R) that satisfy specific conditions (such as peak height ratio, peak spacing, and significance). _double_ring On the other hand, the radius Rmax of the maximum value of I(R) is monitored. When Rmax exhibits a bimodal distribution over time (i.e., two Rmax peaks with similar significance), and this bimodal state persists at least once in time (e.g., occurring at t or t-6h), then an Rmax bimodal event is considered detected. The final event probability P... _ERC (t) can be derived from P _double_ring It is composed of the Rmax bimodal event index. When P _ERC (t) When the preset threshold is exceeded, the system activates the ERC trigger flag I. _ERC (Set to 1).

[0083] Step S402: When the ERC trigger flag is activated, a micro-patch with a cropping scale smaller than the center of the reanalysis meteorological data is cropped around the current location.

[0084] Meteorological data analysis typically involves large-scale data (e.g., a 5°×5° area, approximately 555km×555km). ERC disturbances, however, occur at the typhoon core and are very small in scale. Therefore, when I... _ERC When =1, the system performs a small-scale pruning operation around the current position (e.g., the predicted anchor point p). _t Cut out, for example, 10km×10km or 50km×50km center micro-patterns.

[0085] Step S403: Apply super-resolution reconstruction to the central micro-patch to amplify the perturbation signal and generate a high-resolution micro-patch. Since the central micro-patch is cropped from the coarse-grid reanalysis data, its resolution is insufficient to reveal the structure of the ERC.

[0086] Optionally, a lightweight residual upsampling network is used to perform super-resolution processing on the central micro-patch (e.g., spatial resolution magnification ×2 to ×4) to enhance and amplify the perturbation signal in the micro-patch, generating a high-resolution micro-patch.

[0087] Step S404: Extract ERC micro-perturbation features from the high-resolution micro-pattern. Based on this, the system can calculate a series of features characterizing the ERC physical process. Optionally, the ERC micro-perturbation features include: radial intensity gradient change rate, double-ring morphological index, and eyewall width variation.

[0088] Step S404: Input the ERC micro-perturbation features, mutation-aware 2D representation and geometric orientation encoding into the orientation-preserving low-rank mapping module to generate path deflection increments.

[0089] In some embodiments, mutation-aware and geometrically consistent interactions are performed on 2D and 3D branch features to generate a basic prediction of the typhoon path, and mutation-aware 2D representations and geometric orientation codes are also obtained; and event-conditional micro-perturbation amplification and path coupling are performed, including: when the ERC trigger flag is activated, the ERC micro-perturbation features, mutation-aware 2D representations and geometric orientation codes are input to the direction-preserving low-rank mapping module to generate path deflection increments;

[0090] Specifically, when I _ERC When =1, the system feeds ERC micro-perturbation features, mutation-aware 2D representations, and geometric orientation encodings into a low-rank mapping module (GCA-LRB-ERC) dedicated to orientation preservation for ERC events.

[0091] As an optional implementation, the GCA-LRB-ERC module can share part or all of its structure with the GCA-LRB module. To make this module dedicated to micro-perturbation correction, its parameters can be optimized: the isobaric layer selection (PL-Gate) is forced to Top-K=1, i.e., only the most critical single isobaric layer is selected. Optionally, the ERC micro-perturbation characteristics are used as prior weights α for this layer. _level The algorithm then performs re-normalization; when performing low-rank bilinear interactions, the rank R is optionally constrained (less than R in the interaction module), and a small scaling factor η∈[0.1, 0.5] is applied to the module output. Based on this, the module maps (projects) the high-dimensional ERC micro-perturbation features onto a geometric direction consistent with the path motion, generating a low-dimensional path deflection increment.

[0092] Step S406: Before vector synthesis of the basic typhoon path forecast and the path deflection increment, in other words, before vector synthesis, the path deflection increment needs further processing. The method further includes: calculating the deep-level mean wind from the reanalysis meteorological data to determine the direction of the dominant airflow. The deep-level mean wind is a meteorological indicator used to determine the direction of the main steering airflow of a typhoon. In this embodiment, optionally, at the current position p... _tWithin a 5°×5° region centered on the wind, and across multiple key isobaric layer sets (e.g., {850, 700, 500, 400, 300, 250 hPa}), regional-layer averaging (with different weights applied to different layers, such as 0.4 for lower layers and 0.6 for middle and upper layers) is performed to obtain the average wind vector (U). _DL V _DL The direction d of this vector. _flow This refers to the direction of the dominant airflow.

[0093] Step S407: Apply physical constraints to the path deflection increment to generate the pruned path deflection increment; the physical constraints include: limiting the magnitude of the path deflection increment to not exceed the base prediction (p # _24h The amplitude is set at a preset ratio, and the angle between the path deflection increment and the dominant airflow direction is limited to a preset angle threshold. Since ERC is a micro-perturbation, the resulting path deflection (Δp) is... # _ERC The path deflection increment should not violate the basic physical laws of typhoon motion (e.g., it should not cause the typhoon to be too large or to move against the prevailing airflow). Therefore, the path deflection increment needs to be pruned before synthesis.

[0094] For example, restrict |Δp # _ERC |≤0.2•|p # _24h -p _t That is, the magnitude (amplitude) of the deflection increment should not exceed 20% of the total predicted displacement amplitude of the foundation. Calculate Δp. # _ERC Vector and d _flow The angle between the vectors. This angle must be less than a preset angle threshold θ. _max The symbol |•| represents the Euclidean magnitude of the vector, i.e., its amplitude.

[0095] Optionally, the θ _max It is not a fixed value, but based on training set statistics (e.g., taking the 95th percentile θ of the angle deviation within the mutation window). _p95 This value is empirically approximately 60°, and can be adaptive, for example, based on the current mutation probability P. _mutation (t) Fine-tune. Only the path deflection increment after being clipped by the above constraints can proceed to the next step.

[0096] Step S408: The basic typhoon path prediction generated through geometrically consistent interaction is vector-synthesized with the path deflection increment (i.e., the trimmed path deflection increment) to obtain the final typhoon path prediction. If the ERC trigger flag is not activated, the final typhoon path prediction is made equal to the basic typhoon path prediction. In other words, the trimmed path deflection increment is used to replace the original path deflection increment for vector synthesis. That is, final prediction = basic prediction + trimmed path deflection increment. Simultaneously, the prediction uncertainty (i.e., the prediction uncertainty σ of the model's output for the final prediction) is updated. _fina ).

[0097] Specifically, if ERC is not triggered, then σ _final The prediction uncertainty σ equals the basic prediction _24h If ERC is triggered, σ _final It is the updated uncertainty (e.g., the prediction uncertainty σ of the basic prediction). _24h The variance is approximated by superimposing the variances of the uncertainty introduced by the ERC perturbation.

[0098] It should be understood that this step constitutes the two-stage prediction model of this invention: basic prediction + event correction. When I _ERC When the value is 0, the path deflection increment after pruning is a zero vector, and the final prediction automatically backs down to the base prediction, allowing the model to run smoothly.

[0099] Example 5 describes the specific implementation steps of the geometry-mutation joint training and leakage prevention inference strategy, and explains the training and inference strategy of the model of the present invention. This strategy jointly optimizes geometric accuracy and mutation response, so as to make the training and inference process consistent and prevent data leakage.

[0100] In some embodiments, based on the target location in the final prediction and the standardized sample pair, geometric-mutation joint training is performed, the total loss is calculated, and the model parameters are updated and applied to the next round of prediction.

[0101] In this embodiment, the specific implementation is as follows:

[0102] Step S501: Estimate the mutation probability P based on the mutation feature vector. _mutation (t); Obtain the final predicted typhoon path and the target location Y in the standardized sample pairs. _t Based on the great circle distance between the final prediction and the target location, the geometric principal loss L is calculated. _geo .

[0103] Wherein, the mutation probability P _mutation(t) (or a specific step) is obtained based on the aberration eigenvector estimation. The geometric principal loss is the main objective of model optimization. Optionally, the great circle distance (Haversine formula) is used to calculate the final predicted and actual geographic distance (in kilometers) of the target location on the Earth's surface.

[0104] Step S502: Construct an exponential mutation weighting coefficient based on the mutation probability; use the exponential mutation weighting coefficient to weight the geometric principal loss to generate a mutation weighted loss term.

[0105] In this step, because path mutation samples are sparse in the total training set (imbalanced sample distribution), if all samples have the same loss weight, the model will tend to optimize smooth paths and ignore the fitting of mutation samples. Therefore, this embodiment constructs an exponential mutation weighting coefficient, for example... Where α is a hyperparameter (e.g., α∈[1,3]). The mutation-weighted loss term L _mut The calculation is as follows:

[0106] .

[0107] It can be seen that when the mutation probability P _mutation When λ(t) is very high (e.g., 0.9), λ(t) will become very large (e.g., exp(2•0.9)≈6.05), and the loss L for that sample will be much higher. _geo The mutation probability is amplified; however, when the mutation probability is very low (e.g., 0.1), λ(t) approaches 1 (e.g., exp(2•0.1)≈1.22), and the loss weight is close to the baseline. This causes the optimizer to focus on and correct samples identified as mutations.

[0108] Based on this, specific meteorological events are monitored, and the micro-perturbation amplification and path coupling of the basic forecast based on the typhoon path are performed to further generate event-conditioning loss during training; and the training further includes:

[0109] Step S503: Read the current position from the final predicted typhoon path and the standardized sample pair; calculate the deviation of the predicted heading relative to the historical trend and the deviation of the predicted speed relative to the historical trend based on the final prediction, and generate a geometric smoothing regularization term. This step (L) _smooth As a regularization term, it is used to ensure that the predicted path is geometrically smooth and physically reliable, avoiding abrupt prediction results that violate the inertia of motion.

[0110] Specifically, the system calculates and predicts the heading. _pred and speed _pred (i.e., from the current position p) _tThe vector pointing to the final prediction is compared with the expected heading and expected velocity extrapolated from historical position sequences. This regularization term can optionally be in the form of a squared term:

[0111]

[0112] Among them, Δheading _pred and Δspeed _pred These are the deviations in heading and speed, respectively, and β1 and β2 are weighting coefficients.

[0113] Step S504: Incorporate the mutation-weighted loss term and the geometric smoothing regularization term into the total loss; apply the total loss to perform parameter updates and produce the trained model parameters.

[0114] Event Conditioning Loss L _ERC (or L) _event ) is produced during training. When I _ERC When =1, this loss term (e.g.) ) is activated to specifically constrain prediction accuracy under ERC events; when I _ERC When =0, L _ERC =0.

[0115] Final total loss L _total The aggregated loss terms can be expressed as the following formula:

[0116] Among them, L _mut Already L _geo The weighted terms.

[0117] In some implementations, the total loss may be Among them, L _mut It is λ(t)•L _geo The total loss is further as follows:

[0118] Among them, L _mut It is an independent item that is included in the total loss.

[0119] As an optional training strategy, to further handle sample imbalance and uncertainty, the total loss L _total It can also be multiplied by the overall sample weight w _sample The w _sample It can be determined by focal-like weights w _focal (Handling mutations - smoothing imbalances); mutation weight λ(t) (amplifying mutation loss); prediction uncertainty based on estimation σ _final Calculated weights (Reducing the penalty for samples with high uncertainty) is a joint decision made by three parts.

[0120] The function `stop_grad` is an operation performed during model training (backpropagation phase) to prevent gradients from flowing through its internal parameters. In other words, σ _final The value was used to calculate w _σ (which in turn affects the total loss), but the gradient for updating the model parameters will not pass through w. _σ Feedback to the calculation of σ _final The network branch. The uncertainty σ _final It serves only as a scalar for loss weighting, not as an optimization objective, to stabilize the training process. ε is a numerical stability constant, taking the value 1e-8, to prevent division-by-zero errors, and is used for normalization and probability calculations.

[0121] Based on this, optimizers such as AdamW are used, according to L _total Backpropagate the gradient and perform parameter updates.

[0122] Step S505: During the training phase, anti-leakage training scheduling is adopted, switching between the real anchor point and the predicted anchor point of the previous time step according to probability, and selecting the pruning anchor point. If the real anchor point is used 100% during training, the model will not be able to learn how to correct its own (predicted anchor point) error, resulting in performance collapse during inference (when 100% of the predicted anchor point is used).

[0123] Optionally, the switching probability p (the probability of using the real anchor) is not a fixed value, but decreases dynamically with training (epoch) (e.g., decays from 0.9 to 0.5), allowing the model to gradually adapt to using its own predicted anchor in the later stages of training.

[0124] Step S506: Calculate the deep mean wind from the reanalysis meteorological data to determine the prevailing airflow direction; apply anisotropic Gaussian jitter to the local tangent plane of the clipped anchor point to generate jittered anchor points; the anisotropic Gaussian jitter has a parallel standard deviation along the prevailing airflow direction and an orthogonal standard deviation perpendicular to the prevailing airflow direction, and the parallel standard deviation and the orthogonal standard deviation are not equal. It should be understood that this step is a data augmentation strategy, executed after selecting the anchor point (whether a real anchor point or a predicted anchor point).

[0125] Because typhoon positioning itself involves observational errors, and the uncertainty of a typhoon's path in the direction of the prevailing airflow (which is typically greater) is anisotropic compared to the uncertainty perpendicular to the airflow direction (which is typically smaller), the jitter applied in this embodiment is also anisotropic.

[0126] Accordingly, the dominant airflow direction d is calculated. _flow On the local tangent plane of the anchor point, along d _flow Directional sampling jitter δ _parallel (Standard deviation is σ) _parallel ), and at the same time in vertical d_flow Directional sampling jitter δ _perp (Standard deviation is σ) _perp ), where σ _parallel With σ _perp They are not equal.

[0127] Optionally, (e.g. σ) _parallel =0.15°, σ _perp =0.05°), simulating greater uncertainty along the airflow direction. This jitter (δ) _parallel δ _perp Rotate and apply to the anchor point to generate a jittery anchor point.

[0128] Step S507: Perform anchor pruning using dithered anchors to produce standardized sample pairs for asymmetric multi-source coding. The method employs a consistent anti-leakage anchor pruning mechanism. During training, pruning anchors switch between the real anchor and the predicted anchor from the previous time step according to scheduling probability. In the prediction method, the rolling pruning step uses the predicted anchor from the previous time step, ensuring consistency between training and inference and preventing data leakage.

[0129] Anchor pruning is essentially rolling pruning. During training, pruning is performed using a jittered anchor as the center. A consistent anti-leakage anchor pruning mechanism summarizes the training and inference strategies. For example, during training, anchors are switched between the true anchor and the predicted anchor from the previous time step with a decaying probability p, applying anisotropic jitter. During inference, rolling pruning is performed (100%) using the predicted anchor from the previous time step.

[0130] Optionally, during inference, jitter can be applied (e.g., performing multiple jitter samplings and integrating the predictions) to obtain more robust predictions or assess prediction uncertainty. Furthermore, to ensure geographic consistency, longitude unrolling and wraparound clipping can be optionally performed during anchor point clipping, whether in training or inference, to correctly handle typhoon paths crossing the ±180° International Date Line.

[0131] As another possible implementation of the training process, the system may also include an early stopping criterion. In addition to monitoring the validation set MDE, alternatively, mutation window subset error and ERC trigger subset error should also be specifically monitored to ensure that the model's performance does not degrade on critical sparse samples.

[0132] Example 6 provides a specific and reproducible computational case to illustrate the computational process and technical effects of calculating mutation features (heading rotation) and loss weighting. Accordingly, it demonstrates how the angle rotation normalization function avoids numerical jumps caused by bearing rotation when calculating kinematic features (such as the rate of change of the heading rate of change upon which the rate of change of heading rate of change depends). Furthermore, it demonstrates how the exponential mutation weighting coefficient amplifies the loss of sparse mutation samples using the estimated mutation probability. In this invention, calculating higher-order kinematic features (such as the rate of change of heading and the rate of change of heading rate of change) is crucial for identifying path mutations.

[0133] Assume the typhoon is undergoing a dramatic northward turn across the 0° / 360° azimuth line. At time t-Δ (e.g., t-6h), its heading... _t-Δ The heading is 350° (northwest). At time t+Δ (e.g., t+6h), its heading... _t+Δ The angle is 10° (northeast). The time interval 2Δ for calculating the symmetric difference is 12 hours.

[0134] If we use a simple arithmetic difference to calculate the time difference Δheading:

[0135]

[0136] Based on this difference, the rate of change of course is calculated. _heading ):

[0137]

[0138] It can be seen that the calculated result of -28.33° / h is a large numerical value and a physically incorrect turning rate, wrongly identifying a smooth right turn (clockwise) of 20° as a sharp left turn (counterclockwise) of 340°. The rate of change of heading calculated based on this erroneous value will mislead the abrupt change feature vector.

[0139] In calculating this time difference, this invention applies an angle wrap normalization function (wrap_π), which normalizes the time difference (-340°) to a preset angle range, such as (-180°, 180°). The normalized heading time difference is as follows:

[0140]

[0141] The normalization function correctly identified that the shortest path from 350° to 10° is a 20° clockwise (positive) turn. The rate of change of heading is then calculated based on this normalized difference.

[0142]

[0143] The calculated result of +1.67° / h is a physically reasonable and accurate turning rate.

[0144] By comparing +1.67° / h (in this invention) with -28.33° / h (comparison), it can be seen that applying the angle-around normalization function is a necessary step to calculate the accurate and robust rate of change of heading and its rate of change, and is a technical prerequisite for realizing mutation sensing.

[0145] In the training strategy of this invention, the loss of mutated samples needs to be weighted. Assuming the model is being trained, α is an exponentially weighted hyperparameter, optionally α = 2.0. Sample A (mutated sample) corresponds to the sample with the high rate of change of heading calculated above. The system estimates its mutation probability P. _mutation (t) is very high, for example, P _mutation (t) = 0.9.

[0146] Sample B (smooth sample) corresponds to a sample with a smooth path. The system estimates its mutation probability P. _mutation (t) is very low, for example, P _mutation (t) = 0.1. Assume that the model produces the same geometric principal loss for the final prediction of the target location for both samples, for example, 80km.

[0147] The exponential mutation weighting coefficient λ(t) is constructed based on the mutation probability.

[0148] λ(t) = exp(α•P) _mutation (t));

[0149] For sample A (mutated sample), λ(A) = exp(2.0•0.9) = exp(1.8) ≈ 6.0496.

[0150] Next, the geometric principal loss is weighted using exponential mutation weighting coefficients to generate the mutation-weighted loss term L. _mut L _mut (A) = λ(A)•L _geo =6.0496•80km≈483.97km;

[0151] For sample B (smooth sample), λ(B) = exp(2.0•0.1) = exp(0.2) ≈ 1.2214; calculate its mutation-weighted loss term L. _mut :L _mut (B) = λ(B)•L geo =1.2214•80km≈97.71km; Compare with L _mut (A) (483.97km) and L _mut (B) (97.71km).

[0152] It can be seen that although the original geometric errors of the two samples are the same (both 80km), the training strategy of this invention makes the L generated by the mutant sample A... _mut The loss signal strength is nearly 5 times that of the smoothed sample B (483.97 / 97.71≈4.95). This demonstrates how the mutation-weighted loss term acts as an amplifier in the total loss function, enabling the optimizer to preferentially and more forcefully correct for high-P signals identified as mutations during backpropagation. _mutation The sparse samples of (t) enable a key enhancement of the ability to predict mutation paths.

[0153] Example 7: A high-precision intelligent prediction system for typhoon paths based on a multi-source asymmetric fusion network and its hardware implementation are provided, which can be used to execute the methods of any one of Examples 1 to 6 above.

[0154] Accordingly, the system includes: a processor; a memory storing computer-executable instructions; when the instructions are executed by the processor, they cause the system to perform a method as described in any one of embodiments one through six above, or a combination thereof.

[0155] Specifically, the system can be implemented as one or more servers, cloud computing platforms, or embedded meteorological devices. Memory (e.g., non-volatile memory such as hard disks or SSDs, or volatile memory such as RAM) is used to store the following data:

[0156] Historical typhoon track data and reanalysis meteorological data.

[0157] The computer-executable instructions of the aforementioned methods collectively constitute, for example, an asymmetric multi-source coding module, a mutation-aware and geometrically consistent interaction (MAGI) module, an event-conditioning (MPAN-Geo) module, and a joint training and inference module.

[0158] The trained model parameters produced by training.

[0159] The processor (e.g., CPU, GPU, or dedicated AI acceleration chip such as TPU) is configured as follows:

[0160] During the training phase, the training dataset and instructions are loaded, and geometric principal loss calculation, mutation weighted loss calculation, smoothing regularization term calculation, event loss incorporation and total loss aggregation are performed. Anti-leakage anchor pruning and jittering are performed, and the trained model parameters are produced through backpropagation.

[0161] During the inference phase, the trained model parameters are loaded, and the rolling prediction process is strictly followed. Rolling pruning (using the prediction anchor point of the previous time step) is performed; encoding and basic prediction generation are performed (calling the MAGI module); event monitoring and path coupling are performed (calling the MPAN-Geo module) to generate the final prediction; the final prediction is used as the prediction anchor point for the next iteration, and it is incorporated into the inference trajectory sequence to output the typhoon path prediction result.

[0162] In some alternative implementations, the system may also include a data acquisition module for acquiring the latest track data and reanalysis meteorological data from meteorological data centers (such as CMA, ECMWF) in real time or near real time, supporting operational rolling forecasts.

[0163] As an alternative implementation, Mutation-Aware and Geometrically Consistent Interaction (MAGI) can also be:

[0164] Accordingly, the location sequence (covering t-12h, t-11h, ..., t, where t is the current time) and the reanalysis meteorological tensor are read, and the velocity v is calculated on the location sequence according to spherical geometry. _t The calculations include: heading, rate of change of heading, and rate of change of the rate of change of heading; simultaneously, the calculations of vertical wind shear (wind vector difference at 850 hPa and 250 hPa) and geopotential height gradient rate of change in the reanalysis field within the isobaric layer. In some embodiments, the symmetric time difference of the geopotential height at 500 hPa isobaric layer can also be calculated to obtain the geopotential height time difference. The symmetric differences of the above quantities within the window [t-6h, t, t+6h] form abrupt change feature vectors.

[0165] Optionally, the mutation feature vector is read, and the original probability p0 is obtained by passing it through two fully connected layers-normalization-activation layers (hidden dimension 64 (hidden)); the mutation probability P is obtained by temperature scaling and threshold smoothing. _mutation (t); To alleviate sample imbalance, focal-like weights w are also calculated. _focal =(1-p0) γ (γ is set to 2) as a loss-weighted reference to obtain the mutation probability P. _mutation (t) and weight w _focal .

[0166] Optionally, read the mutation probability P _mutation (t), according to the routing function w _mut =sigmoid(κ•(P _mutation (t)-τ _mut )) and w _smooth =1-w _mut Generate two-stream fusion weights (κ is the slope, τ is the weights). _mut (As a threshold); the weights are subjected to a 3-step moving average over the time axis to suppress jitter, resulting in the stream fusion weight w. _mutwith w _smooth .

[0167] Optionally, scalar time series such as trajectory and intensity in standardized sample pairs are read, fed into a two-layer Bi-LSTM (hidden dimension h2=256, time window T2=12 steps), and the combination of the last time step and attention weighting is taken to obtain a smooth flow representation.

[0168] Optionally, read the mutation feature vector f _t With route weight w _mut For scalar time series such as trajectories, a mutation gate g is introduced outside the input gate of the LSTM. _mut =sigmoid(W _m •f _t +b _m ); and the candidate states c of the LSTM t ~ Use g _mut Magnification can be described by the following formula: c _t =f _gate ⊙c _t -1+i _gate ⊙(g _mut ⊙c t ~ The output is normalized and the residuals are used to obtain the mutation flow representation. A specialized mechanism is implemented to enhance memory updates when mutations occur.

[0169] Optionally, read the 3D branch features F obtained after 3D encoding of the reanalysis meteorological tensor. _3D With mutation feature vector f _t Calculate the score z for the isobaric layer level∈{1000, 850, 500, 250hPa}. _level =qᵀ•[pool(F _3D_level ), f _t ], α is obtained by softmax (temperature τ∈[0.5,2]). _level Retain the Top-K layers (K∈{1,2}) and perform expectation pooling to obtain a sparse 3D representation and layer weights α. _level If an ERC is subsequently triggered, the event signal will force K=1.

[0170] Optionally, read position p _t With position sequence, p _t Mapped to the geocentric unit vector e _t The unit tangent vector τ is obtained by using the great circle differential. _t With the tangential velocity vector v _t The two are concatenated and normalized to form a geometric direction code d. _t =[τ _t v _t], thus obtaining the geometric direction code d _t .

[0171] Optionally, read the smoothed stream representation F _2D_s Mutation-sensing 2D representation F _2D_m Sparse 3D representation, geometric orientation encoding and stream fusion weights w _mut w _smooth Press F _2D_m =w _mut •F _2D_m +w _smooth •F _2D_s ; Formation of mutation-sensing 2D representation F _2D_mixed Next, perform the low-rank geometric outer product, i.e.:

[0172] B(x, y) = Σ _r=1..R (U) _r •x,τ _t • (V) _r •y,v _t );

[0173] Where R∈[4, 8], U _r V _r The dimensional parameters are shared in the isobaric layer; a geometrically consistent representation is obtained by applying signed-sqrt and L2 normalization to B. This is used as the main input to the regression head.

[0174] Optionally, the geometrically consistent representation and the low-order aligned representation are read, and the predicted longitude and latitude are output through the multilayer perceptron and the geographic coordinate regression head. The basic prediction is obtained through spherical projection. The prediction uncertainty is output by the branch head (for subsequent loss weighting or confidence interval calculation).

[0175] In other alternative implementations, event-conditional ERC perturbation amplification and path coupling (MPAN-Geo) can also be achieved in the following ways:

[0176] Specifically, the reanalysis meteorological tensor and the current location are read, and the radial maximum wind speed radius Rmax and eyewall intensity distribution are calculated within a radius of 50–150 km. When an increase in the probability of a double-ring structure is detected or the evolution of Rmax shows a double peak and meets the threshold condition, the ERC trigger flag I is output. _ERC ∈{0,1} and the probability of the event.

[0177] in I _ERCWhen =1, the current location is read as the anchor point. A 10km×10km central micro-patch is cropped under geographic coordinates, and a small-range equidistant projection is made from the Earth's surface to the local tangent plane to obtain the central micro-patch (multiple isobaric layers, multiple variables). The central micro-patch is read and passed through a lightweight residual upsampling network (magnification ×2 to ×4, 32 channels and 4 traces) to improve spatial resolution and contrast, resulting in a high-resolution micro-patch.

[0178] Furthermore, high-resolution micro-patterns are read, and the radial intensity gradient change rate, double-ring morphology index, and eyewall width change are calculated; these are then concatenated with the mutation feature vector to form the ERC micro-perturbation feature. The ERC micro-perturbation feature and layer weight α are then read. _level Force Top-K=1 and use ERC micro-perturbation features as prior pairs α _level Perform a re-normalization to obtain a single-layer weight and a single-layer 3D representation that are consistent with the event.

[0179] Furthermore, reading mutation-aware 2D representations F _2D_mixed Single-layer 3D representation and geometric orientation encoding d _t By employing a low-rank geometric outer product but restricting the rank to R, a small scaling factor η∈[0.1, 0.5] is applied to the output channel to obtain the path deflection increment (in vector form).

[0180] Furthermore, read the path deflection increment Δp # _ERC Compared with the basic prediction p # _24h Limit the deflection amplitude and direction according to physical consistency, for example, limit |Δp # _ERC |≤0.2•|p # _24h -p _t |and the angle between it and the direction of the prevailing airflow is ≤θ _max (θ) _max (A value of 60° can be used) to obtain the path deflection increment after clipping. Read the base prediction and the path deflection increment after clipping, and synthesize the vectors to obtain the final prediction = base prediction + path deflection increment after clipping. At the same time, update the prediction uncertainty (i.e. the prediction uncertainty of the model for the final prediction output, such as the variance superposition approximation).

[0181] Furthermore, read the final prediction p # _final With target position Y _t and ERC trigger flag I _ERC With the probability of event P _ERC (t), when I _ERC When =1, construct the event conditional loss L. _ERC =γ•Haversine(p #_final Y _t •P _ERC (t), γ∈[0.5, 2]; when I _ERC =0 when L _ERC =0, and the event loss is obtained.

[0182] In some embodiments, an exemplary calculation scheme for the mutation feature vector is as follows:

[0183] Read the position sequence (covering t-6h, t, t+6h) and define the central difference on a sphere. Let Δ = 6 hours, given the great circle distance dist between two times a and b. _gc (a, b) (unit: km) and the spherical azimuth angle bearing (a, b) (unit: radians). Then the tangential velocity vector v is... _t =dist _gc (p) _t-Δ p _t+Δ ) / (2Δ)。

[0184] heading _t =bearing(p _t-Δ p _t+Δ ).

[0185] Acceleration of heading _heading_t ≈wrap_π (heading) _t+Δ -heading _t-Δ ) / (2Δ)。

[0186] rate of change of heading (jerk) _t ≈wrap_π(acceleration) _heading_t+Δ -acceleration _heading_t-Δ ) / (2Δ)。

[0187] It should be understood that wrap_π() normalizes the angle difference to (-π, π], avoiding jumps caused by longitude ±180° and azimuth rotation. This yields the symmetric difference feature g. _sym (t) = {v _t ,heading _t ,leration _heading_t jerk _t}

[0188] Furthermore, read g _sym (t) and the reanalysis meteorological tensor, g _sym The components of (t) are expanded according to sin / cos (for heading, acceleration). _heading(jerk), and with scaling (z-score, parameters from the training set), we obtain the vector f. _motion (t). Calculate the vector wind difference modulus (vertical wind shear) at 850 hPa and 250 hPa, calculate the geopotential height time difference dZ_dt (t±6h symmetric difference) at 500 hPa, and calculate the circumferential mean and variance of the relative vorticity ζ850 at 850 hPa, forming f. _env (t). Mutation feature vector f _t =concat(f _motion (t), f _env (t) is standardized according to the mean / variance of the training set to obtain the mutation feature vector.

[0189] Based on this, the missing detection mask and mutation feature vector are read, and nearest neighbor spatiotemporal interpolation is used for the missing detection dependencies; dimensions that cannot be interpolated are filled with 0 and a missing detection indicator bit is appended before the probability header to obtain the robust f. _t_stable It can be used directly to replace the mutation feature vector in subsequent steps.

[0190] In other embodiments, the implementation method of low-rank decomposition and isobaric layer dimension sharing further includes:

[0191] Accordingly, read batch features x _batch (From F) _2D_mixed ) and y _batch (From sparse 3D representation or single-layer 3D representation), initialize U _r With V _r The process involves the following two stages:

[0192] Cold start (without statistical priors), for dimension d _x With d _y Sampling Gaussian matrices, performing QR decomposition, and obtaining the orthogonal basis from the first R columns, let U = U0•σ _u V = V0•σ _v , where σ _u =1 / sqrt(d _x ), σ _v =1 / sqrt(d _y This yields the orthogonal scaling initializations U and V. U0 and V0 are the initial orthogonal frames of the corresponding dimensional spaces.

[0193] A warm start (if there is prior knowledge, it is recommended to execute it once at the end of the first epoch) is performed to calculate the cross-modal covariance C. _xy =E[xy T (Batch or sliding window), for C _xy Perform singular value decomposition C _xy =PΣQ T Take the first R singular vectors, and let U = P _R ΣR 1 / 2 V=Q _R Σ _R 1 / 2 The learning rate is reduced to 0.2 of its original value for stable training in the first 1–3 epochs, resulting in U and V for SVD warm-up. Here, P, Σ, and Q are SVD decomposition matrices used for GCA-LRB parameter warm-start initialization. P is the d _x The matrix is ​​a k×k diagonal singular value matrix (left singular vector matrix), Σ is a k×k diagonal singular value matrix, and Q is a d _y ×k matrix (right singular vector matrix) (d _x For the dimension of the input variable, d _y (where k is the number of singular values, representing the dimension of the output variable).

[0194] Furthermore, read the layer-by-layer decomposition 3D representation {y _l} and layer weight α _l (From PL-Gate), two equivalent implementations are provided, choose one:

[0195] Method 1 (sharing before aggregation), perform gated aggregation. * =Σ _l α _l •y _l , for y * Using the same group {V _r}. This is equivalent to all layers sharing a single V. _r Among them, y * This refers to the comprehensive features after gating aggregation.

[0196] Method 2 (aggregated and shared with layered scale), perform v on each layer separately. _l，r =(V _r •y _l v _t ), with α _l Aggregate b _r =Σ _l α _l •v _l，r At this time, U _r V _r In layer-dimensional sharing, while α _l As a learnable or event-driven sparse gate; optional layer-rank scales s _l，r (Scalar), but the default value is s to control complexity. _l，r =1.

[0197] Both methods maintain the same set of U _r V _r Layer-level sharing meets the requirements for shared parameters.

[0198] In some other embodiments, the ERC double-ring structure and the Rmax double-peak determination specifically include:

[0199] Correspondingly, read the wind speed field of the central micro-patch or the 850 hPa layer, and calculate the tangential wind component V in polar coordinates for the radius R ∈ [10 km, 100 km] and the azimuth θ ∈ [0, 2π). _tan (R, θ). Form the radius-averaged profile I(R) = mean _θ V _tan (R, θ). Find local maxima R1 < R2 on I(R) that satisfy the two-peak interval ΔR ∈ [10 km, 60 km], the peak height ratio I(R2) / I(R1) ∈ [0.6, 1.1], and the prominence of both peaks exceeds the threshold P _min .

[0200] Calculate the double-ring index BI = min(prominence1, prominence2) • (ΔR / 60 km) • min(I(R2) / I(R1), I(R1) / I(R2)). Obtain the double-ring probability P through logistic mapping _double_ring = sigmoid(a • BI + b) (a and b are fitted from the training set). Here, prominence is the peak significance, which is the height difference between the peak height and the nearest saddle point and is used to identify real peaks.

[0201] Optionally, read I(R), and define Rmax as the radius where the maximum value of I(R) is located; when there are more than two significant peaks, take the two largest peaks Rmax1 and Rmax2. If |Rmax2 - Rmax1| ∈ [10 km, 60 km] and I(Rmax2) / I(Rmax1) ∈ [0.6, 1.1], mark it as an Rmax double-peak; to avoid transient false positives, it is required to last at least once in the adjacent one (t ± 6 h); combined with P _double_ring , form the event probability P _ERC (t) = 0.5 • P _double_ring + 0.5 • I{Rmax double-peak}. Output I _ERC (threshold trigger).

[0202] Read the ERC micro-perturbation characteristics and P _double_ring , P _ERC (t), splice them and then output the event prior vector through a lightweight calibration head for single-layer selection and mapping scaling.

[0203] In some other embodiments, the physical constraints and the dominant airflow direction can also be:

[0204] Further, read the reanalysis meteorological tensor at the position p _tCalculate the deep mean wind on a 5°×5° region centered at the center, using the isobaric layer set {850, 700, 500, 400, 300, 250 hPa}: U _DL =mean _区域，层 U, V _DL =mean _区域，层 V (weights of 0.4 and 0.6 can be applied to lower and middle-upper levels); dominant airflow direction d _flow =atan2(V _DL U _DL (radians), converted to unit tangent vector. This yields the dominant airflow direction vector d. _flow It is used in deflection limiting and jitter orientation.

[0205] Optionally, the training set trajectory and deep wind direction are read, the angular deviation distribution of the predicted relative dominant airflow within the abrupt change window is statistically analyzed, and its 95th quantile is taken as θ. _p95 Where θ _p95 ≈p95 (empirical default, as a conservative upper bound); execute the adaptive policy, θ _max =clamp(θ) _p95 (45°, 75°), can be finely adjusted according to the mutation intensity to obtain θ. _max =θ _max •(1+0.2•(P _mutation (t)-0.5)); used for physical cutting.

[0206] Optionally, read Δp # _ERC With d _flow θ _max , will Δp # _ERC Decomposed into components along the dominant airflow and vertical components (approximately linear in the spherical tangent plane); Limit 1: |Δp # _ERC |ERC ball•|p # _24h -p _t |. Limiting 2: Angle (Δp) # _ERC d _flow )fl _max Limiter 3 (optional): If strong convection parameters (such as ζ850 or wind shear) are below the threshold, scale by an additional 0.5. Output the clipped path deflection increment.

[0207] The optional embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solution of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A typhoon path high-precision intelligent prediction method based on a multi-source asymmetric fusion network, characterized in that, The method comprises the following steps: acquiring typhoon historical path data and reanalysis meteorological data, performing rolling clipping based on the predicted anchor point at the previous moment, and outputting a standardized sample pair at the current moment; performing asymmetric multi-source coding on the standardized sample pair to generate 2D branch features and 3D branch features, and performing mutation perception and geometric consistent interaction thereon to generate a basic prediction of the typhoon path; monitoring a specific meteorological event, performing event-conditioned micro-disturbance amplification and path coupling based on the basic prediction of the typhoon path to generate a final prediction of the typhoon path; taking the final prediction of the typhoon path as the predicted anchor point at the next moment, and integrating it into the reasoning trajectory sequence; wherein the predicted anchor point refers to the typhoon center position output by the model when performing the prediction task at the previous moment; wherein monitoring a specific meteorological event further comprises: monitoring an eye wall replacement precursor using the standardized sample pair at the current moment to generate an ERC trigger flag; when the ERC trigger flag is activated, a center micro patch with a scale smaller than the reanalysis meteorological data is cropped around the current position; a super-resolution reconstruction amplification disturbance signal is applied to the center micro patch to generate a high-resolution micro patch; and ERC micro-disturbance features are extracted from the high-resolution micro patch; wherein the mutation perception and geometric consistent interaction are performed on the 2D branch features and the 3D branch features to generate the basic prediction of the typhoon path; and the event-conditioned micro-disturbance amplification and path coupling specifically comprises: when the ERC trigger flag is activated, the ERC micro-disturbance features, the mutation perception 2D representation and the geometric direction encoding are input into a direction-preserving low-rank mapping module to generate a path deflection increment; the basic prediction of the typhoon path generated through the geometric consistent interaction and the path deflection increment are vector synthesized to obtain the final prediction of the typhoon path; if the ERC trigger flag is not activated, the final prediction of the typhoon path is equal to the basic prediction of the typhoon path; wherein the mutation perception and geometric consistent interaction performed on the 2D branch features and the 3D branch features comprises: based on the standardized sample pair at the current moment, kinematic features and environmental prior features are calculated and combined into a mutation feature vector, wherein the kinematic features include the rate of change of the rate of change of the heading change rate calculated based on the spherical geometric symmetric difference, and the environmental prior features include the vertical wind shear and the potential height time difference; the 2D branch features and the mutation feature vector are input into a mutation flow encoder; the mutation flow encoder introduces a mutation gate outside the input gate of the recurrent neural network unit; the mutation gate is controlled by the mutation feature vector to dynamically amplify the update strength of the candidate state inside the unit, and outputs a mutation flow representation; The mutation perception and geometric consistent interaction on the 2D branch feature and the 3D branch feature further comprises: applying an isobaric layer selection gate to the 3D branch feature to release Top-K key layers to generate a sparse 3D representation; based on the position sequence in the standardized sample pair, constructing a geometric direction code on the spherical tangent plane, the geometric direction code containing a unit tangent vector and a tangent velocity vector; feeding the mutation flow representation, the sparse 3D representation and the geometric direction code into a direction-preserving low-rank bilinear interaction module; the interaction module projects the mutation flow representation to the unit tangent vector direction and projects the sparse 3D representation to the tangent velocity vector direction to output a geometric consistent representation; and generating a basic prediction of the typhoon path based on the geometric consistent representation.

2. The method of claim 1, wherein, Before vector composition of the basic prediction of the typhoon path and the path deflection increment, the method further comprises: calculating the deep layer mean wind from the reanalysis meteorological data to determine the dominant airflow direction; applying physical constraints to the path deflection increment to generate a clipped path deflection increment; the physical constraints include: limiting the amplitude of the path deflection increment to be not more than a preset proportion of the amplitude of the basic prediction, and limiting the included angle between the path deflection increment and the dominant airflow direction to be not more than a preset angle threshold; using the clipped path deflection increment to replace the path deflection increment for vector composition.

3. The method of claim 1, wherein, Based on the spherical geometric symmetric difference, the rate of change of the rate of change of the heading change rate is calculated, further comprising: when calculating the heading change rate and the rate of change of the heading change rate, applying an angle wrap normalization function to the time difference of the heading and the time difference of the heading change rate; the angle wrap normalization function normalizes the time difference to a preset angle interval to avoid numerical jumps caused by azimuth wrap; the rate of change of the heading change rate is calculated using the normalized time difference.

4. The method of claim 1, wherein, The asymmetric multi-source coding on the standardized sample pair comprises: applying a multi-scale 3D convolution to the 3D reanalysis meteorological tensor in the standardized sample pair to perform causal time compression, compressing its time dimension to a shorter summary dimension to generate 3D branch features; and the asymmetric multi-source coding further outputs a low-order alignment representation, comprising: applying a linear projection of the first order to the summaries of the 2D branch features and the 3D branch features to retain the common scale across modalities and suppress the noise dimension; wherein the summary of the 2D branch feature refers to reducing the 2D branch feature in time series to a vector; and the summary of the 3D branch feature refers to reducing the 3D branch feature to a vector.

5. A typhoon path high-precision intelligent prediction system based on a multi-source asymmetric fusion network, characterized in that, comprising: a processor; a memory, the memory having computer executable instructions stored thereon; the instructions, when executed by the processor, cause the system to perform the method of any one of claims 1 to 4.

Images