Typhoon track prediction method and system fusing physical constraints and path mutation recognition

Abstract

The application provides a typhoon path prediction method and system fusing physical constraints and path mutation identification. The method comprises the following steps: obtaining a multidimensional input feature tensor through a feature tensor generation module; performing time modeling through a multiscale modeling module to obtain a deep feature tensor, performing significance guidance through a significance guidance module to obtain a prediction path sequence; identifying a bending angle through a bending event identification module to obtain a plurality of path anomaly scores, introducing a plurality of preset disturbances through a counter disturbance analysis module to determine path confidence, and outputting a path prediction result through a path prediction output module. According to the technical scheme of the embodiment of the application, the typhoon path can be finely modeled, the PINNs physical constraint module is used to ensure that the prediction result meets the physical conservation principle, the bending event identification is used to enhance the path mutation detection capability, the disturbance analysis is used to realize the uncertainty evaluation, and the typhoon path prediction result is more accurate, reliable and physically consistent.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent weather prediction, and in particular to a typhoon path prediction method and system fusing physical constraints and path mutation recognition. BACKGROUND

[0002] With the intensification of global climate change, the uncertainty of typhoon intensity and path is increasing, and high-precision typhoon path prediction is a key technical problem in current meteorological services and disaster warning systems. Traditional path prediction methods mainly include statistical regression models (such as the CLIPER series), numerical weather prediction models (such as ECMWF, WRF, etc.), and intelligent prediction methods based on machine learning. Among them, although the numerical model has a physical basis, it still has problems such as large error and high computational resource consumption in the subtropical high control area and path mutation segment; while the deep learning method, with its advantages in big data fitting and pattern recognition, has been gradually applied to path prediction, intensity estimation and other directions in recent years, showing good potential.

[0003] However, the existing typhoon path prediction methods lack physical consistency, cannot effectively identify typhoon path mutations, and the uncertainty of the prediction process is difficult to quantify, making it difficult to guarantee the reliability of typhoon prediction. SUMMARY

[0004] The present application aims to at least solve one of the technical problems existing in the prior art. To this end, the present application proposes a typhoon path prediction method and system fusing physical constraints and path mutation recognition, which can balance physical consistency, mutation warning capability and uncertainty evaluation capability, and improve the reliability of typhoon path prediction.

[0005] In a first aspect, an embodiment of the present application provides a typhoon path prediction method fusing physical constraints and path mutation recognition, applied to a typhoon path prediction system, the typhoon path prediction system comprising a feature tensor generation module, a multi-scale modeling module, a saliency guiding module, a PINNs physical constraint module, a bending event recognition module, an adversarial perturbation analysis module and a path prediction output module, the PINNs physical constraint module being used to assist in supervising the training of the multi-scale modeling module and the saliency guiding module, and the method comprising:

[0006] typhoon information of a target typhoon is input into the feature tensor generation module to obtain a multi-dimensional input feature tensor;

[0007] the input feature tensor is input into the multi-scale modeling module for time modeling to obtain a deep feature tensor, and the deep feature tensor is guided for saliency by the saliency guiding module to obtain a prediction path sequence, wherein the prediction path sequence comprises predicted positions at multiple prediction times;

[0008] Based on the predicted path sequence, a plurality of path anomaly scores are identified by the bending event identification module through the bending angle, a path confidence is determined by introducing a plurality of preset disturbances through the adversarial perturbation analysis module, and a path prediction result is output through the path prediction output module, wherein the path prediction result includes the path confidence of the predicted path sequence and the path anomaly score of each predicted time.

[0009] According to some embodiments of the application, the typhoon information of the target typhoon is input into the feature tensor generation module to obtain a multi-dimensional input feature tensor, including:

[0010] A plurality of typhoon variables are obtained from the typhoon information, wherein the typhoon variables include typhoon wind speed, typhoon pressure, typhoon temperature, typhoon potential height, and typhoon specific humidity;

[0011] A preset collection time step, a preset resolution, and a plurality of preset atmospheric levels along the vertical direction are obtained, wherein the preset resolution is used to divide a plurality of spatial horizontal grids;

[0012] The collection time step, the preset resolution, the preset atmospheric level, and the typhoon variable are input into the feature tensor generation module to obtain the input feature tensor.

[0013] According to some embodiments of the application, the multi-scale modeling module includes a CNN network and a time modeling network, and the input feature tensor is input into the multi-scale modeling module for time modeling to obtain a deep feature tensor, including:

[0014] The input feature tensor is input into the CNN network to extract a spatio-temporal feature sequence, wherein the spatio-temporal feature sequence includes spatio-temporal features corresponding to a plurality of collection time points, the spatio-temporal features include local spatial vortexes and wind field structures, and the collection time points are determined based on the collection time step;

[0015] The spatio-temporal feature sequence is input into the time modeling network, and the deep feature tensor is output by the time modeling network based on the global dependency between a plurality of the spatio-temporal features.

[0016] According to some embodiments of the application, the deep feature tensor is guided for significance by the significance guiding module to obtain a predicted path sequence, including:

[0017] A plurality of historical typhoon paths are obtained, and a significance map is constructed after normalizing a plurality of the historical typhoon paths;

[0018] The deep feature tensor and the significance map are fused according to a preset significance guiding factor to obtain a fused feature tensor, and the predicted path sequence is predicted based on the fused feature tensor.

[0019] The calculation formula of the fusion feature tensor is: wherein, is the fusion feature tensor, is the saliency guide factor, is the deep feature tensor, is the saliency map, represents point-by-point multiplication.

[0020] According to some embodiments of the present application, the PINNs physical constraint module is preset with a water vapor conservation constraint, a momentum conservation constraint and an energy conservation constraint, and the multi-scale modeling module and the saliency guide module are obtained based on the PINNs physical constraint module for auxiliary supervision training, comprising:

[0021] A water vapor conservation loss function is constructed based on the water vapor conservation constraint, wherein the formula of the water vapor conservation loss function is , is the water vapor conservation loss function, is the water vapor budget calculated by the model at the path point, is the true water vapor budget;

[0022] A momentum conservation loss function is constructed based on the momentum conservation constraint, wherein the formula of the momentum conservation loss function is , is the momentum conservation loss function, is the movement vector of the path point predicted by the model, is the background environmental wind speed;

[0023] An energy conservation loss function is constructed based on the energy conservation constraint, wherein the formula of the energy conservation loss function is: , is the energy conservation loss function, is the total energy of a unit atmospheric column predicted by the model, is the actual total energy of a unit atmospheric column, and satisfies , is the total energy of a unit atmospheric column, is the specific heat capacity at constant pressure, is the typhoon temperature, is the acceleration of gravity, is the typhoon geopotential height, is the horizontal component of the typhoon wind speed in the east-west direction, is the horizontal component of the typhoon wind speed in the north-south direction;

[0024] The total loss function is determined as a loss function in a training process of the multi-scale modeling module and the saliency guiding module based on weighted summation of the water vapor conservation loss function, the momentum conservation loss function and the energy conservation loss function.

[0025] According to some embodiments of the present application, the bending event identification module comprises a 3-layer fully connected network and a softmax layer, the softmax layer comprising a plurality of preset bending events, a plurality of path anomaly scores being obtained by identifying bending angles through the bending event identification module, comprising:

[0026] A plurality of continuous predicted bending angles are determined based on a plurality of continuous predicted positions;

[0027] When a plurality of continuous predicted bending angles are all greater than a preset angle value, a path composed of corresponding predicted positions is determined as a mutant path;

[0028] The mutant path is sequentially input into the 3-layer fully connected network and then into the softmax layer to determine a bending event classification probability, the path anomaly score being obtained by normalizing the bending event classification probability, and the path anomaly score being associated with each predicted position of the mutant path.

[0029] According to some embodiments of the present application, the preset disturbance comprises air pressure disturbance and temperature disturbance, and a path confidence is determined by introducing a plurality of preset disturbances through the adversarial disturbance analysis module, comprising:

[0030] A plurality of groups of air pressure disturbances and temperature disturbances are introduced in initial meteorological field data determined based on the typhoon information to generate a plurality of groups of disturbance samples;

[0031] Based on any disturbance sample, a plurality of disturbance paths are predicted according to a plurality of predicted positions, and the path confidence is determined according to a mean value and a standard deviation of the plurality of disturbance paths.

[0032] According to some embodiments of the present application, a path prediction result is output through the path prediction output module, comprising:

[0033] The predicted time corresponding to each path anomaly score is determined based on the predicted position;

[0034] The path confidence is associated with the predicted path sequence;

[0035] The path prediction result is visually displayed, wherein the path anomaly score and the predicted position are sequentially displayed based on the predicted time.

[0036] In a second aspect, an embodiment of the present application provides a typhoon path prediction system fusing physical constraints and path mutation recognition, comprising:

[0037] a feature tensor generation module configured to generate a multi-dimensional input feature tensor according to typhoon information of a target typhoon;

[0038] a multi-scale modeling module configured to perform time modeling according to the input feature tensor to obtain a deep feature tensor;

[0039] a saliency guidance module configured to perform saliency guidance on the deep feature tensor to obtain a predicted path sequence, wherein the predicted path sequence comprises predicted positions at multiple prediction time points;

[0040] a PINNs physical constraint module configured to assist in supervised training of the multi-scale modeling module and the saliency guidance module;

[0041] a bending event recognition module configured to recognize bending angles according to the predicted path sequence to obtain multiple path anomaly scores;

[0042] an adversarial perturbation analysis module configured to determine path confidence by introducing multiple preset perturbations in the predicted path sequence;

[0043] a path prediction output module configured to output a path prediction result according to the predicted path sequence, the path anomaly scores and the path confidence, wherein the path prediction result comprises the path confidence of the predicted path sequence and the path anomaly score at each prediction time point.

[0044] In a third aspect, an embodiment of the present application provides a typhoon path prediction system fusing physical constraints and path mutation recognition, comprising at least one control processor and a memory connected in communication with the at least one control processor; the memory stores instructions executable by the at least one control processor, and the instructions are executed by the at least one control processor to enable the at least one control processor to perform the typhoon path prediction method fusing physical constraints and path mutation recognition as described in the first aspect.

[0045] According to the typhoon path prediction method fusing physical constraints and path mutation recognition provided in the embodiments of the present application, the typhoon information of a target typhoon is input into the feature tensor generation module to obtain a multi-dimensional input feature tensor; the input feature tensor is input into the multi-scale modeling module to perform time modeling to obtain a deep feature tensor, and the deep feature tensor is guided for significance by the significance guiding module to obtain a predicted path sequence, wherein the predicted path sequence includes predicted positions at multiple prediction times; based on the predicted path sequence, a bending angle is identified by the bending event identification module to obtain multiple path anomaly scores, a path confidence is determined by introducing multiple preset disturbances by the adversarial perturbation analysis module, and a path prediction result is output by the path prediction output module, wherein the path prediction result includes the path confidence of the predicted path sequence and the path anomaly score of each prediction time. According to the technical solution of the embodiments of the present application, the typhoon path can be finely modeled, the PINNs physical constraint module is used to ensure that the prediction result meets the physical conservation principle, the bending event identification is used to enhance the path mutation detection capability, the disturbance analysis is used to realize the uncertainty evaluation, and thus a more accurate, reliable and physically consistent typhoon path prediction result is output. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 FIG. 1 is a schematic diagram of a typhoon path prediction system provided in an embodiment of the present application;

[0047] Figure 2 FIG. 2 is a flowchart of a typhoon path prediction method fusing physical constraints and path mutation recognition provided in another embodiment of the present application;

[0048] Figure 3 FIG. 3 is a functional schematic diagram of each module in a typhoon path prediction system provided in another embodiment of the present application;

[0049] Figure 4 FIG. 4 is a structural diagram of a typhoon path prediction system fusing physical constraints and path mutation recognition provided in another embodiment of the present application. DETAILED DESCRIPTION

[0050] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the drawings, in which the same or similar notations represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the drawings are exemplary and are only used to explain the present application, and cannot be understood as a limitation of the present application.

[0051] In the description of the present application, it should be understood that the orientation description, such as the orientation or position relationship indicated by the upper, lower, front, rear, left, right and the like, is based on the orientation or position relationship shown in the drawings, and 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.

[0052] In the description of the present application, one or more is understood as one or more, more than two is understood as more than two, greater than, less than, more than, etc. are understood as not including the number, above, below, etc. are understood as including the number. If the first, second is described, it is only for the purpose of distinguishing technical features, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated or the order of technical features indicated.

[0053] In the description of the present application, unless otherwise explicitly limited, the words such as setting, installing, connecting and the like should be broadly understood, and those skilled in the art can reasonably determine the specific meaning of the above words in the present application in combination with the specific content of the technical solution.

[0054] The embodiment of the present application provides a typhoon path prediction method and system fusing physical constraints and path mutation recognition, and the typhoon path prediction method fusing physical constraints and path mutation recognition comprises: inputting typhoon information of a target typhoon into the feature tensor generation module to obtain a multi-dimensional input feature tensor; inputting the input feature tensor into the multi-scale modeling module for time modeling to obtain a deep feature tensor, and performing significance guidance on the deep feature tensor through the significance guidance module to obtain a prediction path sequence, wherein the prediction path sequence comprises predicted positions at multiple prediction times; based on the prediction path sequence, a plurality of path anomaly scores are obtained by identifying a bending angle through the bending event identification module, a path confidence is determined by introducing a plurality of preset disturbances through the adversarial disturbance analysis module, and a path prediction result is output through the path prediction output module, wherein the path prediction result comprises the path confidence of the prediction path sequence and the path anomaly score of each prediction time. According to the technical scheme of the embodiment of the present application, the typhoon path can be finely modeled, the PINNs physical constraint module is used to ensure that the prediction result meets the physical conservation principle, the bending event identification is used to enhance the path mutation detection capability, and the disturbance analysis is used to realize the uncertainty evaluation, so that a more accurate, reliable and physically consistent typhoon path prediction result is output.

[0055] First, the professional terms involved in the present embodiment are explained as follows:

[0056] Convolutional Neural Networks (CNNs) are deep learning models designed specifically for processing grid-structured data (such as images and videos). Their core feature is that they significantly reduce the number of parameters through local connections and weight sharing, while preserving the spatial structure information of the data. They have become a core technology in the field of computer vision.

[0057] Physics-Informed Neural Networks (PINNs) are a machine learning framework that integrates physical laws with deep learning. The core idea is to embed physical laws (such as partial differential equations and conservation laws) as constraints into the training of the neural network, thereby enabling the modeling and prediction of physical systems. By incorporating physical constraints such as partial differential equations into the neural network's loss function, the physical consistency and interpretability of the model are improved.

[0058] BendNet, a bend event classifier, is a path anomaly identification model based on machine learning or meteorological rules. It is used to detect and classify abrupt changes in typhoon paths, such as sharp turns, path loops, or interactions between two typhoons, providing support for abrupt change early warning.

[0059] Adversarial perturbation analysis refers to artificially adding small but structured perturbations (such as local wind field or air pressure changes) to the initial field data of a typhoon to test the sensitivity of the prediction model to input uncertainties, thereby evaluating its stability and robustness.

[0060] Saliency guidance map: A spatial guidance map constructed based on the frequency of occurrence of historical paths, used to enhance the model's focus on common path regions.

[0061] Momentum conservation: In typhoon track prediction, the model output should maintain momentum consistency with changes in the environmental wind field. Momentum conservation can serve as a physical constraint, reducing unreasonable deviations between predictions and actual meteorological evolution.

[0062] Uncertainty Quantification (UQ) refers to characterizing the reliability of prediction results through probabilistic modeling. Applied to typhoon track prediction, it can generate prediction intervals and confidence probability distributions, which can be used for risk classification, disaster response strategy formulation, and other scenarios.

[0063] The Fujiwhara effect refers to the phenomenon where two tropical cyclones, about 1,000 kilometers (1,400 kilometers) apart, interact with each other through vortices and rotate around a common center, eventually merging or weakening. This phenomenon was proposed by Japanese meteorologist Sakuhei Fujiwara.

[0064] First, refer to Figure 1 , Figure 1This is a schematic diagram of a typhoon path prediction system provided in an embodiment of the present invention. The prediction system in this embodiment includes a feature tensor generation module, a multi-scale modeling module, a saliency guidance module, a PINNs physical constraint module, a bending event identification module, an adversarial disturbance analysis module, and a path prediction output module.

[0065] It should be noted that, as Figure 1 As shown, the typhoon path prediction system in this embodiment connects the feature tensor generation module, multi-scale modeling module, saliency guidance module, PINNs physical constraint module, bending event recognition module, adversarial disturbance analysis module, and path prediction output module in sequence. The PINNs physical constraint module is located in the middle of the overall training chain, receiving feature inputs from the multi-scale modeling module and the saliency guidance module, and constructing a loss function using physical conservation equations (mass, water vapor, momentum, energy) for auxiliary supervision. Since the training utilizes the physical constraints of the PINNs physical constraint module, the features output from the saliency guidance module to the bending event recognition module and the adversarial disturbance analysis module are also affected by the physical constraints. This allows the PINNs physical constraint module to indirectly influence the bending event recognition module and the adversarial disturbance analysis module, thereby outputting more accurate, reliable, and physically consistent typhoon path prediction results. The PINNs physical constraint module mainly provides supervision of the loss function during training; therefore, it does not participate in data processing in actual applications and will not be described again later.

[0066] The following is based on the appendix Figure 1 The prediction system shown further illustrates the technical solution of the embodiments of the present invention.

[0067] Reference Figure 2 , Figure 2 This is a flowchart illustrating a typhoon path prediction method that integrates physical constraints and path abrupt change identification, provided by an embodiment of the present invention. This typhoon path prediction method includes, but is not limited to, steps S10, S20, and S30. Specifically:

[0068] S10: Input the typhoon information of the target typhoon into the feature tensor generation module to obtain a multi-dimensional input feature tensor.

[0069] It should be noted that the typhoon information in this embodiment can be obtained from the reanalysis data (e.g., ERA5) or observation data of the target typhoon. The typhoon information may include the typhoon's wind speed, air pressure, and temperature, etc., and the specific variables can be selected according to actual needs. It is worth noting that the typhoon information is continuous data over a period of time, so the information can be collected according to the sequence of data collected at the set collection time steps.

[0070] It should be noted that, as Figure 3As shown, the typhoon path prediction system in this embodiment also includes a data input module and a data preprocessing module. The data input module is used to input the acquired typhoon information, and the data preprocessing module is used to preprocess the variables of the typhoon information. Data preprocessing is a technology well known to those skilled in the art, and will not be described in detail here.

[0071] It should be noted that after the feature tensor generation module obtains the typhoon information, it constructs a multi-dimensional input feature tensor based on multiple variables of the typhoon information, providing a spatiotemporal multi-dimensional feature foundation for the prediction of the typhoon path.

[0072] S20, the input feature tensor is input to the multi-scale modeling module for time modeling to obtain a deep feature tensor. The deep feature tensor is then guided by the saliency guidance module to obtain a prediction path sequence, which includes the prediction positions at multiple prediction times.

[0073] It should be noted that in this embodiment, the multi-scale modeling module first extracts spatiotemporal features from the input feature tensor and performs time modeling. Since typhoon information is collected based on the acquisition time step, time modeling can capture global dependencies across time steps. Finally, the deep feature representation corresponding to the time step of the multi-scale modeling module is output. In this embodiment, this deep feature representation is used as the deep feature tensor. The multi-scale modeling module can perform preliminary modeling of the typhoon path. Therefore, the deep feature tensor is the initially predicted typhoon path.

[0074] It should be noted that typhoon paths have significant statistical distribution characteristics. Therefore, in this embodiment, the deep feature tensor is guided by a saliency guidance module, so that historical saliency features are fused with the deep feature tensor. The fused features are used as the predicted path sequence obtained in this case. The predicted path sequence includes the predicted positions at multiple prediction times.

[0075] It should be noted that the multi-scale modeling module and the saliency guidance module are obtained through supervised training based on the PINNs physical constraint module, which can help avoid non-physical jumps or unreasonable deviations from the path and improve the physical consistency of the prediction.

[0076] S30: Based on the predicted path sequence, the bending angle is identified by the bending event identification module to obtain multiple path anomaly scores. Multiple preset perturbations are introduced by the adversarial perturbation analysis module to determine the path confidence. The path prediction output module outputs the path prediction results, which include the path confidence of the predicted path sequence and the path anomaly score at each prediction time.

[0077] It should be noted that this embodiment adds a bending event recognition module to the typhoon path prediction system. The bending event recognition module identifies the bending angle of the typhoon path. When a bending angle exists, sudden events such as sharp turns, loops, and Fujiwhara effects may occur. Bending event recognition can be achieved using the bending event classifier BendNet, or other classifiers can be used. As long as the bending angle is determined based on the preset path sequence, it can effectively capture abnormal changes such as sharp turns and loops in the typhoon path.

[0078] It should be noted that this embodiment also includes an anti-disturbance analysis module. When the predicted path sequence is already obtained, a preset disturbance is introduced through this module. This disturbance alters the typhoon's meteorological field data, resulting in a disturbed path different from the original path. If the difference between the disturbed path and the original path is significant, the uncertainty of the original path is high, and the path confidence is poor. Conversely, if the difference is small, the uncertainty of the original path is low, and the path confidence is high. By applying disturbances to the initial field and guiding the path distribution, an uncertainty assessment mechanism is formed. The output confidence interval provides quantitative support for path reliability, facilitating the hierarchical processing of forecast information by the operational system.

[0079] It should be noted that this embodiment integrates the predicted path sequence, path anomaly score, and path confidence score through the path prediction output module, so that the final output includes the predicted path sequence, the path anomaly score and path confidence score at each prediction time, and can output the forecast map in a visual form, so that the typhoon path prediction results can be displayed intuitively. While improving physical consistency, it further reflects the identified bending events and path confidence scores in the path prediction results, thereby improving the reliability of typhoon prediction.

[0080] In another embodiment, step S10 specifically includes, but is not limited to, the following steps:

[0081] S11. Multiple typhoon variables are obtained from typhoon information, including typhoon wind speed, typhoon air pressure, typhoon temperature, typhoon geopotential height, and typhoon specific humidity.

[0082] S12, acquire the preset acquisition time step, preset resolution and multiple preset atmospheric layers along the vertical direction, wherein the preset resolution is used to divide multiple spatial horizontal grids;

[0083] S13: Input the acquisition time step, preset resolution, preset atmospheric level and typhoon variables into the feature tensor generation module to obtain the input feature tensor.

[0084] It should be noted that the typhoon information in this embodiment can be obtained from multi-source typhoon variables from reanalysis data (such as ERA5) or observational data. The typhoon variables in this embodiment include typhoon wind speed, typhoon air pressure p (hPa), typhoon temperature T (K), typhoon geopotential height Z (gpm), and typhoon specific humidity q. Among them, the typhoon wind speed includes the horizontal velocity component along the east-west direction. and the horizontal velocity component along the north-south direction .

[0085] It should be noted that this embodiment presets the data acquisition time step T. For example, setting T=12 ensures that data is collected once per hour over the past 12 hours. The preset atmospheric levels are multiple typical atmospheric levels in the vertical direction, such as selecting five typical atmospheric levels: 850hPa, 700hPa, 500hPa, 300hPa, and 200hPa. The preset resolution is the horizontal resolution. For example, setting H×W=64×64, where H is the grid length and W is the grid width, divides the space into a 64×64 horizontal grid with a resolution of 1 degree.

[0086] It should be noted that, as Figure 3 As shown, after determining the above parameters, they are input into the feature tensor generation module to obtain the input feature tensor. The shape of the input feature tensor is determined by the above parameters. For example, the input feature tensor is (T=12, L=5, H=64, W=64, F=6), where T is the acquisition time step, F is the number of typhoon variables, L is the number of preset atmospheric layers, H is the grid length, and W is the grid width. This tensor serves as the input to the subsequent multi-scale modeling module, providing a spatiotemporal multidimensional feature basis for typhoon path prediction.

[0087] In another embodiment, in step S20, the multi-scale modeling module includes a CNN network and a temporal modeling network. The input feature tensor is input to the multi-scale modeling module for temporal modeling to obtain a deep feature tensor, which specifically includes, but is not limited to, the following steps:

[0088] S211, Input the input feature tensor into the CNN network to extract the spatiotemporal feature sequence, wherein the spatiotemporal feature sequence includes spatiotemporal features corresponding to multiple acquisition times, and the spatiotemporal features include local spatial vortices and wind field structures, and the acquisition time is determined based on the acquisition time step;

[0089] S212 inputs the spatiotemporal feature sequence into the temporal modeling network, and outputs a deep feature tensor based on the global dependencies between multiple spatiotemporal features through the temporal modeling network.

[0090] It should be noted that, as Figure 3As shown, the multi-scale modeling module includes a CNN network and a temporal modeling network. The CNN network can use 3×3×3 convolutional kernels to extract local spatial vortex and wind field structures as spatiotemporal feature sequences. The temporal modeling network can be an LSTM (128 hidden units), a Transformer (4 layers, 8 attention heads), a GRU, or a BiLSTM, etc. After obtaining the high-dimensional input tensor, the model first uses a CNN neural network to extract local spatial vortex features and wind field structures. Subsequently, the extracted spatiotemporal feature sequences are input into a Transformer-based temporal modeling network (containing 4 layers, 8 attention heads per layer) to capture global dependencies across time steps. Finally, this module outputs a deep feature representation with a predicted time step length of 24. ,in This indicates the embedded feature dimension. The output will be passed to the saliency guidance module and the PINNs physical constraint module for further processing.

[0091] In another embodiment, in step S20, the deep feature tensor is saliency-guided by the saliency guidance module to obtain the predicted path sequence, which specifically includes, but is not limited to, the following steps:

[0092] S221: Obtain multiple historical typhoon paths, normalize the multiple historical typhoon paths, and construct a saliency map;

[0093] S222, the deep feature tensor and the saliency map are fused according to the preset saliency guiding factor to obtain the fused feature tensor, and the predicted path sequence is predicted based on the fused feature tensor;

[0094] The formula for calculating the fusion feature tensor is as follows: ,in, To fuse feature tensors, As a significant leading factor, For deep feature tensors, For saliency plot, It represents point-by-point multiplication.

[0095] It should be noted that typhoon paths exhibit significant statistical distribution characteristics. The typhoon path prediction system in this embodiment constructs a saliency map M based on historical typhoon path data from the past 30 years and normalizes it to [0,1]. This saliency map is fused with a weighted factor and a deep feature tensor output from multi-scale modeling. The fused feature tensor is... , As a significance guide factor (usually taken as 0.5–1.0), the fused feature tensor can maintain the model's ability to respond to abnormal path regions while ensuring that the model pays attention to common path regions.

[0096] In another embodiment, the PINNs physical constraint module is pre-set with water vapor conservation constraints, momentum conservation constraints, and energy conservation constraints. The multi-scale modeling module and the saliency guidance module are obtained based on the PINNs physical constraint module for assisted supervised training, specifically including but not limited to the following steps:

[0097] S01, Construct a water vapor conservation loss function based on the water vapor conservation constraint, wherein the formula for the water vapor conservation loss function is: , Let the water vapor conservation loss function be... The water vapor budget is calculated for the predicted path points in the model. For the actual water vapor balance;

[0098] S02, Construct a momentum conservation loss function based on the momentum conservation constraint, wherein the formula for the momentum conservation loss function is: , The momentum conservation loss function, The path point movement vectors predicted by the model. Background wind speed;

[0099] S03, Construct an energy conservation loss function based on energy conservation constraints, where the formula for the energy conservation loss function is: , The energy conservation loss function is... This represents the total energy per unit atmospheric column predicted by the model. This represents the actual total energy per unit atmospheric column, and satisfies... , This refers to the total energy per unit atmospheric column. For isobaric specific heat capacity, The temperature of the typhoon. It is the acceleration due to gravity. The typhoon's position and height. This represents the horizontal component of the typhoon's wind speed in the east-west direction. This represents the horizontal component of the typhoon's wind speed in the north-south direction.

[0100] S04. The total loss function is obtained by weighted summation of the water vapor conservation loss function, momentum conservation loss function and energy conservation loss function. The total loss function is then used as the loss function in the training process of the multi-scale modeling module and the saliency guidance module.

[0101] It should be noted that, under the constraint of water vapor conservation, the water vapor flux is defined as: ,in Let Q be the water vapor flux and Q be the specific humidity of the air. Let be the wind speed. Perform divergence calculations on it to obtain the water vapor budget. The water vapor conservation loss is: ,in, The water vapor budget is calculated for the predicted path points of the model. : The actual water vapor balance derived from reanalysis data.

[0102] It should be noted that the momentum conservation constraint requires that the typhoon's motion trend be consistent with the mid-level environmental wind field, that is... ,in The path point movement vectors predicted by the model. The background wind speed is 500 hPa or 700 hPa.

[0103] It should be noted that, under the constraint of energy conservation, the total energy per unit atmospheric column consists of three parts: thermal energy, potential energy, and kinetic energy, expressed as follows: ,in Specific heat capacity at constant pressure ( ), The temperature of the typhoon. It is the acceleration due to gravity. The typhoon's position and height. This represents the horizontal component of the typhoon's wind speed in the east-west direction. For the horizontal component of typhoon wind speed in the north-south direction, the energy conservation loss is defined as: This constraint is used to avoid unreasonable energy spikes or drops during the prediction process.

[0104] It should be noted that, after determining the above three constraints, the total loss function in this embodiment is a weighted sum of the above three loss functions, such that the trajectory coordinate error and the physical consistency error together constitute the training objective. The expression for the total loss function is: ,in To predict the mean squared error (MSE) of each path in the path sequence. These are different preset hyperparameter weights.

[0105] It should be noted that, as Figure 1 and Figure 3 As shown, the PINNs physical constraint module in Figure 1 Positioned in the middle, it acts as a bridge between the front-end feature output and subsequent path prediction. Instead of being a standalone forward computation module, it uses the total loss function to inversely constrain the entire network during training. Located in the middle of the overall training chain, it receives feature inputs from the multi-scale modeling module and the saliency guidance module, and uses physical conservation equations (mass, water vapor, momentum, energy) to construct a loss function for auxiliary supervision. Simultaneously, its result affects "bend event recognition," "adversarial disturbance analysis," and the final "path prediction output." This ensures that the output data and constraints during actual prediction are not isolated but rather progressive and interactive, resulting in more accurate, reliable, and physically consistent typhoon path predictions.

[0106] In another embodiment, the bending event recognition module includes a 3-layer fully connected network and a softmax layer. The softmax layer includes multiple preset bending events. In step S30, the bending angle is identified by the bending event recognition module to obtain multiple path anomaly scores, which specifically includes, but is not limited to, the following steps:

[0107] S311, determine multiple consecutive predicted bending angles based on multiple consecutive predicted locations;

[0108] S312, when multiple consecutive predicted bending angles are all greater than the preset angle value, the path formed by the corresponding predicted positions is determined as the mutation path;

[0109] S313, the mutation path is passed through 3 fully connected network layers and then input into the softmax layer to determine the classification probability of the bending event. The classification probability of the bending event is normalized to obtain the path anomaly score. The path anomaly score is associated with each predicted position of the mutation path.

[0110] It should be noted that the bending angle in this embodiment is determined by at least three consecutive predicted positions in time. The predicted bending angle is defined as follows: , Let be the predicted position at time t. Meanwhile, during typhoon prediction, because it is a continuous process over time, new predicted deflection angles are continuously generated as the value of t increases. If only one predicted deflection angle is greater than the preset angle value (30 degrees), it may be an isolated case. Therefore, this embodiment needs to ensure that the predicted deflection angles persist for a certain period of time before the path formed by these predicted positions can be determined as a sudden road surface change. For example... If the duration is longer than 2 hours, it is marked as a path change. In this embodiment, the duration is determined by the number of predicted bend angles. For example, a predicted bend angle is determined every 10 minutes. Then, 12 consecutive predicted bend angles that are greater than the preset angle value are sufficient.

[0111] It should be noted that, as Figure 3 As shown, this embodiment employs a 3-layer fully connected neural network (64–32–2 nodes) in the bending event recognition module, using softmax to output a path anomaly score, expressed as follows: The output probability value serves as the path anomaly score and is input into the prediction output module. ( ) represents the first fully connected layer. ( ) represents the second fully connected layer. ( ) represents the third fully connected layer. The Softmax layer can classify different bending events, including sharp turns, path loops, or interactions between two typhoons.

[0112] In another embodiment, the preset disturbances include pressure disturbances and temperature disturbances. In step S30, multiple preset disturbances are introduced through the anti-disturbance analysis module to determine the path confidence, specifically including but not limited to the following steps:

[0113] S321, based on typhoon information, determines the initial meteorological field data, introduces multiple sets of pressure disturbances and temperature disturbances into the initial meteorological field data, and generates multiple sets of disturbance samples;

[0114] S322: Based on any perturbation sample, multiple perturbation paths are predicted according to multiple predicted locations, and the path confidence is determined based on the mean and standard deviation of the multiple perturbation paths.

[0115] It should be noted that, given the availability of typhoon information, determining the initial meteorological field data is a technique well-known to those skilled in the art, and will not be elaborated upon here.

[0116] It should be noted that, to quantify the uncertainty of the prediction results, this embodiment incorporates multiple sets of pressure and temperature disturbances into the initial meteorological field data. Pressure disturbances can be at the ±1 hPa or ±0.5 K level, and temperature disturbances can be at the ±0.5 K level. After introducing the disturbances, multiple disturbance samples are determined from the obtained field data. The inference process is repeated 20 times for each set of disturbance samples to obtain multiple disturbance samples. Each disturbance sample is then independently predicted to generate multiple disturbance paths, resulting in a path set. Finally, the mean and standard deviation of the path set are calculated to form a 95% confidence interval as the path confidence level.

[0117] In another embodiment, in step S30, the path prediction output module outputs the path prediction result, which specifically includes, but is not limited to, the following steps:

[0118] S331, Determine the prediction time corresponding to the anomaly score of each path based on the predicted location;

[0119] S332, associates the path confidence with the predicted path sequence;

[0120] S333, visualizes the path prediction results, where path anomaly scores and predicted locations are displayed in the order of prediction times.

[0121] It should be noted that the expression for the predicted path sequence obtained in this embodiment is based on... For example, Let be the latitude at time t. Let t be the longitude and time step. (24 hours). When performing bend event recognition, it is based on the predicted position. Each predicted position can be used as a matching basis to determine the corresponding path anomaly score, thereby associating the path anomaly score with each prediction time. Similarly, the path confidence can also be associated with the corresponding predicted path sequence, so that the output of the predicted path output module not only records the predicted path sequence, but also includes the path anomaly score and path confidence at each prediction time.

[0122] In addition, one embodiment of the present invention provides a typhoon path prediction system that integrates physical constraints and path change identification, comprising:

[0123] The feature tensor generation module is used to generate multi-dimensional input feature tensors based on the typhoon information of the target typhoon.

[0124] A multi-scale modeling module is used to perform time modeling based on the input feature tensor to obtain a deep feature tensor;

[0125] A saliency guidance module is used to saliency guide the deep feature tensor to obtain a predicted path sequence;

[0126] The PINNs physical constraint module is used to assist in supervised training of the multi-scale modeling module and the saliency guidance module.

[0127] A bend event recognition module is used to identify bend angles based on the predicted path sequence to obtain multiple path anomaly scores.

[0128] An anti-disturbance analysis module is used to introduce multiple preset disturbances into the predicted path sequence to determine the path confidence.

[0129] The path prediction output module is used to output a path prediction result based on the predicted path sequence, the path anomaly score, and the path confidence score, wherein the path prediction result includes the path confidence score of the predicted path sequence and the path anomaly score at each prediction time.

[0130] It should be noted that the principles of each of the above modules can be found in the description of the above method embodiments, and will not be repeated here.

[0131] like Figure 4 As shown, Figure 4 This is a structural diagram of a typhoon path prediction system that integrates physical constraints and path abrupt change identification, provided in one embodiment of the present invention. The present invention also provides a typhoon path prediction system that integrates physical constraints and path abrupt change identification, comprising:

[0132] The processor 401 can be implemented using a general-purpose central processing unit (CPU), microprocessor, application specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0133] The memory 402 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 402 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 402 and called by the processor 401 to execute the typhoon path prediction method integrating physical constraints and path change identification of the embodiments of this application.

[0134] Input / output interface 403 is used to implement information input and output;

[0135] The communication interface 404 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0136] Bus 405 transmits information between various components of the device (e.g., processor 401, memory 402, input / output interface 403, and communication interface 404);

[0137] The processor 401, memory 402, input / output interface 403 and communication interface 404 are connected to each other within the device via bus 405.

[0138] This application also provides an electronic device, including the typhoon path prediction system that integrates physical constraints and path change identification as described above.

[0139] This application embodiment also provides a storage medium, which is a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the above-described typhoon path prediction method that integrates physical constraints and path change identification.

[0140] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof. The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separate, and may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0141] It will be understood by those skilled in the art that all or some of the steps and systems in the methods disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically include computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0142] The above provides a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.

Claims

1. A typhoon track prediction method integrating physical constraints and track change identification, characterized in that, An application is made to a typhoon path prediction system, which includes a feature tensor generation module, a multi-scale modeling module, a saliency guidance module, a PINNs physical constraint module, a bending event recognition module, an adversarial perturbation analysis module, and a path prediction output module. The PINNs physical constraint module is used to assist in supervised training of the multi-scale modeling module and the saliency guidance module. The method includes: The typhoon information of the target typhoon is input into the feature tensor generation module to obtain a multi-dimensional input feature tensor; The input feature tensor is input to the multi-scale modeling module for time modeling to obtain a deep feature tensor. The saliency guidance module guides the deep feature tensor to obtain a prediction path sequence, wherein the prediction path sequence includes the prediction position at multiple prediction times. Based on the predicted path sequence, the bending angle is identified by the bending event identification module to obtain multiple path anomaly scores, multiple preset perturbations are introduced by the adversarial perturbation analysis module to determine the path confidence, and the path prediction output module outputs the path prediction result, wherein the path prediction result includes the path confidence of the predicted path sequence and the path anomaly score at each prediction time. The bending event recognition module includes a 3-layer fully connected network and a softmax layer. The softmax layer includes multiple preset bending events. The bending event recognition module identifies bending angles to obtain multiple path anomaly scores, including: Multiple consecutive predicted bending angles are determined based on multiple consecutive predicted positions; When multiple consecutive predicted bending angles are greater than a preset angle value, the path formed by the corresponding predicted positions is determined as a sudden change path. The mutation path is passed through three layers of the fully connected network and then input into the softmax layer to determine the bend event classification probability. The bend event classification probability is normalized to obtain the path anomaly score, and the path anomaly score is associated with each of the predicted positions of the mutation path.

2. The typhoon path prediction method integrating physical constraints and path abrupt change identification according to claim 1, characterized in that, Typhoon information of the target typhoon is input into the feature tensor generation module to obtain a multi-dimensional input feature tensor, including: Multiple typhoon variables are obtained from the typhoon information, including typhoon wind speed, typhoon air pressure, typhoon temperature, typhoon geopotential height, and typhoon specific humidity. The system acquires a preset acquisition time step, a preset resolution, and multiple preset atmospheric layers along the vertical direction, wherein the preset resolution is used to divide the space into multiple horizontal grids. The acquisition time step, the preset resolution, the preset atmospheric level, and the typhoon variable are input into the feature tensor generation module to obtain the input feature tensor.

3. The typhoon path prediction method integrating physical constraints and path abrupt change identification according to claim 2, characterized in that, The multi-scale modeling module includes a CNN network and a temporal modeling network. The input feature tensor is input into the multi-scale modeling module for temporal modeling to obtain a deep feature tensor, including: The input feature tensor is input into the CNN network to extract a spatiotemporal feature sequence, wherein the spatiotemporal feature sequence includes spatiotemporal features corresponding to multiple acquisition times, and the spatiotemporal features include local spatial vortices and wind field structures, and the acquisition times are determined based on the acquisition time steps; The spatiotemporal feature sequence is input into the temporal modeling network, and the temporal modeling network outputs the deep feature tensor based on the global dependencies between multiple spatiotemporal features.

4. The typhoon path prediction method integrating physical constraints and path abrupt change identification according to claim 3, characterized in that, The prediction path sequence is obtained by saliency guidance of the deep feature tensor through the saliency guidance module, including: Obtain multiple historical typhoon paths, and construct a saliency map after normalizing the multiple historical typhoon paths; The deep feature tensor and the saliency map are fused according to a preset saliency guiding factor to obtain a fused feature tensor, and the predicted path sequence is predicted based on the fused feature tensor. The formula for calculating the fused feature tensor is: ,in, For the fused feature tensor, The significance guiding factor, For the deep feature tensor, For the saliency map, It represents point-by-point multiplication.

5. The typhoon path prediction method integrating physical constraints and path abrupt change identification according to claim 1, characterized in that, The PINNs physical constraint module is pre-set with water vapor conservation constraints, momentum conservation constraints, and energy conservation constraints. The multi-scale modeling module and the saliency guidance module are obtained based on the PINNs physical constraint module for assisted supervised training, including: A water vapor conservation loss function is constructed based on the aforementioned water vapor conservation constraint, wherein the formula for the water vapor conservation loss function is as follows: , Let be the water vapor conservation loss function. The water vapor budget is calculated for the predicted path points in the model. For the actual water vapor balance; A momentum conservation loss function is constructed based on the momentum conservation constraint, wherein the formula for the momentum conservation loss function is: , The momentum conservation loss function, The path point movement vectors predicted by the model. Background wind speed; An energy conservation loss function is constructed based on the energy conservation constraint, wherein the formula for the energy conservation loss function is: , The energy conservation loss function is... This represents the total energy per unit atmospheric column predicted by the model. This represents the actual total energy per unit atmospheric column, and satisfies... , This refers to the total energy per unit atmospheric column. For isobaric specific heat capacity, The temperature of the typhoon. It is the acceleration due to gravity. The typhoon's position and height. This represents the horizontal component of the typhoon's wind speed in the east-west direction. This represents the horizontal component of the typhoon's wind speed in the north-south direction. The total loss function is obtained by weighted summation of the water vapor conservation loss function, the momentum conservation loss function, and the energy conservation loss function. This total loss function is then used as the loss function during the training process of the multi-scale modeling module and the saliency guidance module.

6. The typhoon path prediction method integrating physical constraints and path abrupt change identification according to claim 1, characterized in that, The preset disturbances include pressure disturbances and temperature disturbances. Multiple preset disturbances are introduced through the counter-disturbance analysis module to determine the path confidence, including: Based on the typhoon information, initial meteorological field data is determined, and multiple sets of pressure disturbances and temperature disturbances are introduced into the initial meteorological field data to generate multiple sets of disturbance samples. Based on any of the disturbance samples, multiple disturbance paths are predicted according to multiple predicted locations, and the confidence level of the paths is determined according to the mean and standard deviation of the multiple disturbance paths.

7. The typhoon path prediction method integrating physical constraints and path abrupt change identification according to claim 6, characterized in that, The path prediction output module outputs path prediction results, including: Based on the predicted location, determine the predicted time corresponding to each path anomaly score; Associate the path confidence with the predicted path sequence; The path prediction results are visualized, wherein the path anomaly score and the predicted location are displayed in order based on the prediction time.

8. A typhoon path prediction system integrating physical constraints and path change identification, characterized in that, include: The feature tensor generation module is used to generate multi-dimensional input feature tensors based on the typhoon information of the target typhoon. A multi-scale modeling module is used to perform time modeling based on the input feature tensor to obtain a deep feature tensor; A saliency guidance module is used to saliency guide the deep feature tensor to obtain a prediction path sequence, wherein the prediction path sequence includes the prediction positions at multiple prediction times. The PINNs physical constraint module is used to assist in supervised training of the multi-scale modeling module and the saliency guidance module. A bend event recognition module is used to identify bend angles based on the predicted path sequence to obtain multiple path anomaly scores. An anti-disturbance analysis module is used to introduce multiple preset disturbances into the predicted path sequence to determine the path confidence. A path prediction output module is used to output a path prediction result based on the predicted path sequence, the path anomaly score, and the path confidence score, wherein the path prediction result includes the path confidence score of the predicted path sequence and the path anomaly score at each prediction time. The bending event recognition module includes a 3-layer fully connected network and a softmax layer. The softmax layer includes multiple preset bending events. The bending event recognition module identifies bending angles to obtain multiple path anomaly scores, including: Multiple consecutive predicted bending angles are determined based on multiple consecutive predicted positions; When multiple consecutive predicted bending angles are greater than a preset angle value, the path formed by the corresponding predicted positions is determined as a sudden change path. The mutation path is passed through three layers of the fully connected network and then input into the softmax layer to determine the bend event classification probability. The bend event classification probability is normalized to obtain the path anomaly score, and the path anomaly score is associated with each of the predicted positions of the mutation path.

9. A typhoon path prediction system integrating physical constraints and path change identification, characterized in that, It includes at least one control processor and a memory for communicatively connecting to the at least one control processor; the memory stores instructions executable by the at least one control processor, which, when executed by the at least one control processor, enable the at least one control processor to perform the typhoon path prediction method that integrates physical constraints and path change identification as described in any one of claims 1 to 7.

Images