A rapid enhancement prediction method for typhoons based on the fusion of spatiotemporal sequences and multimodal features.

Abstract

This invention relates to the field of typhoon forecasting technology and discloses a method for rapid typhoon enhancement prediction based on the fusion of spatiotemporal sequences and multimodal features. The method includes: constructing a multimodal spatiotemporal sequence dataset and an auxiliary dataset based on optimal typhoon path data and multi-source satellite observation data; performing dimensionality reduction preprocessing on the high-dimensional multimodal spatiotemporal sequence data using a unified manifold approximation and projection method to generate a one-dimensional time-series embedded representation of the typhoon observation sequence; and inputting the one-dimensional time-series embedded representation and the auxiliary data as independent input channels into a trained typhoon observation network model to predict the probability of rapid typhoon enhancement. The typhoon observation network model is a multimodal spatiotemporal fusion deep learning architecture, whose core consists of a variational attention recurrent neural network, and its hyperparameters are optimized using an improved Harris Eagle optimization algorithm. This invention achieves accurate and robust identification of the rapid typhoon enhancement process.

Description

Technical Field

[0001] This invention relates to the field of typhoon forecasting technology, and in particular to a rapid enhancement forecasting method for typhoons based on the fusion of spatiotemporal sequences and multimodal features. Background Technology

[0002] Typhoons, as one of the most destructive extreme weather disasters globally, often exhibit rapid intensification accompanied by a sudden drop in central pressure and a dramatic increase in wind speed, posing a significant threat to maritime shipping, coastal city safety, and emergency response deployment. Due to the highly complex formation and evolution of typhoons, influenced by multi-source observational data (such as remote sensing cloud images, meteorological satellites, sea surface temperature, pressure fields, and temperature and humidity profiles), accurately predicting whether a typhoon will rapidly intensify in the future presents an extremely high technical challenge.

[0003] Current mainstream methods are mostly based on statistical features or single model structures, which have the following shortcomings: 1. Difficulty in fusing multimodal observation data: Remote sensing images, meteorological parameters, cloud top brightness temperature and other features have different dimensions and scales, making it difficult to model them uniformly; 2. Limited spatiotemporal evolution feature expression: Traditional methods are insufficient in handling the nonlinear dynamics of typhoon evolution over time, especially in accurately identifying key turning points before enhancement; 3. Weak model generalization and optimization capabilities: Conventional neural network structures have problems such as unstable training, strong hyperparameter dependence, and easy getting trapped in local optima.

[0004] Therefore, there is an urgent need for a new prediction method that can integrate multimodal information, has strong time series modeling capabilities, and can adaptively optimize parameter configuration, in order to achieve accurate and robust identification of the rapid intensification process of typhoons. Summary of the Invention

[0005] This invention aims to address the shortcomings of existing rapid typhoon enhancement prediction methods in terms of multimodal data fusion, feature representation, and optimization capabilities, and proposes a rapid typhoon enhancement prediction method based on the fusion of spatiotemporal sequences and multimodal features.

[0006] To achieve the above objectives, the following technical solution is adopted:

[0007] A rapid typhoon enhancement prediction method based on the fusion of spatiotemporal sequences and multimodal features includes the following steps:

[0008] Based on typhoon optimal path data and multi-source satellite observation data, a multimodal spatiotemporal sequence dataset and an auxiliary dataset are constructed.

[0009] The Unified Manifold Approximation and Projection (UMAP) method is used to perform dimensionality reduction preprocessing on high-dimensional multimodal spatiotemporal sequence data, generating a one-dimensional time series embedding representation of typhoon observation sequences;

[0010] The one-dimensional time series embedding representation and auxiliary data are used as independent input channels and input into the trained typhoon observation network model to predict the probability of rapid typhoon intensification.

[0011] The typhoon observation network model is a multimodal spatiotemporal fusion deep learning architecture, whose core consists of a variational attention recurrent neural network (VA-RNN) and whose hyperparameters are optimized by an improved Harris Hawks optimization algorithm (DDHHO).

[0012] Furthermore, the construction steps of the multimodal spatiotemporal sequence dataset are as follows:

[0013] N consecutive typhoon observation sequences were collected at 6-hour intervals. Each sequence includes: a) Remote sensing cloud image sequence: depth feature vectors extracted from remote sensing cloud images using a convolutional neural network pre-trained on a basic cloud image dataset; b) Typhoon intensity sequence: maximum wind speed at the center; c) Upward cloud top brightness temperature sequence: infrared brightness temperature measurements; d) Temperature profile feature sequence: vertical temperature distribution features compressed by wavelet; e) Humidity profile feature sequence: vertical humidity distribution features compressed by wavelet; f) Rapid enhancement markers: binary labels indicating an intensity increase ≥ 7 m / s over the next 12 hours.

[0014] Spatiotemporal alignment of multi-source observation data is achieved using cubic spline interpolation.

[0015] Data augmentation is performed through path simulation and feature perturbation to enhance sample diversity.

[0016] Furthermore, the specific process of the dimensionality reduction preprocessing includes:

[0017] A weighted adjacency graph that preserves local structure is constructed in a high-dimensional space, and sample similarity is calculated based on weighted Euclidean distance and local neighborhood scale parameters. The local neighborhood scale parameters are negatively correlated with the rate of change of air pressure at the typhoon center or the intensity of sudden wind speed changes, which is used to enhance the sensitivity to rapid intensification precursors.

[0018] Define a similarity function for robust anomalies in a low-dimensional space;

[0019] The embedding representation is optimized by minimizing the cross-entropy loss of the similarity difference between high- and low-dimensional spaces, and a one-dimensional time series embedding representation is generated through iterative optimization.

[0020] Furthermore, the training process of the typhoon observation network model employs a recurrent neural network with a variational attention mechanism to process the one-dimensional time series embedding representation and the auxiliary data, specifically including:

[0021] The one-dimensional time series embedding representation and the auxiliary data input multi-head attention mechanism are used to generate query vectors, key vectors and value vectors through linear transformation, which are used to capture the correlation of key regions in different time steps;

[0022] A parallel multi-head structure is adopted to fuse multi-scale features: each attention head captures signals from different physical mechanisms, including features of strong convective cloud regions or abrupt changes in temperature and humidity profiles; the outputs of the multiple heads are spliced ​​and fused across heads through a linear mapping matrix to obtain a fused feature map;

[0023] Perform temporal window compression aggregation on the fused feature map, and generate window aggregated features by global average pooling along the time dimension;

[0024] The window aggregation features are input into a recurrent neural network to model the temporal evolution, and the typhoon energy accumulation process is encoded through hidden states; the hidden state of the last time step is input into a classifier to output the rapid enhancement probability for the next 12 hours.

[0025] Furthermore, the classifier consists of a fully connected layer and a sigmoid activation function, and outputs a probability value; and the probability determination threshold is set to 0.55. When the probability value exceeds 0.55, it is determined to be a fast boost.

[0026] The total loss function for model training consists of a binary classification cross-entropy loss function and a temporal embedding reconstruction loss function.

[0027] Furthermore, the multi-head attention mechanism of the variational attention recurrent neural network includes three types of dedicated attention heads:

[0028] Type I attention heads: Mz in total, only accepting one-dimensional time-series embedded feature window sequences. It is used to capture abrupt changes during the development of typhoons;

[0029] Type II attention heads: Ma heads in total, only receiving auxiliary data feature window sequences. Used to monitor intensity transition signals;

[0030] The third type of attention head: a total of Mc heads, simultaneously receiving a one-dimensional time series embedded feature window sequence. and auxiliary data feature window sequence It is used to capture the cross-pattern of two types of features.

[0031] Furthermore, the total loss function for training the model is:

[0032]

[0033] in: The weighting coefficients for the temporal embedding reconstruction loss range from 0.05 to 0.2; the temporal embedding reconstruction loss function... :

[0034]

[0035] To reconstruct the mapping function and , For dimension , To hide the state dimension, Embedding dimensions for one-dimensional time series; For the fully connected layer mapping function, the hidden state is... Mapped to a one-dimensional time series embedding representation at time step t Same dimensions; For layer normalization function;

[0036] Binary cross-entropy loss for:

[0037] ;

[0038] in, : Real and fast tag enhancement; Model prediction confidence.

[0039] Furthermore, the auxiliary data includes: typhoon life cycle identifier, encoded as a binary scalar; and multi-scale intensity difference parameters, encoded as unique heat vectors.

[0040] The variational attention recurrent neural network dynamically integrates the dual-channel outputs through a gating fusion mechanism:

[0041] ;

[0042] in, This is a dynamic gating coefficient, whose value is adaptively adjusted according to the state of the auxiliary data: when the auxiliary data detects an intensity transition... At that time, the gating coefficient When auxiliary data shows a stable state, .

[0043] Furthermore, hyperparameter optimization is performed using an improved Harris Hawks optimization algorithm, including:

[0044] The population is initialized using Tent chaotic mapping combined with random perturbation;

[0045] Design a nonlinear escape energy function to dynamically regulate the ratio of global exploration to local exploitation;

[0046] When the escape energy E≥1, the Cauchy mutation mechanism is used to update the parameters for global exploration and optimization, thereby enhancing the ability to escape local optima.

[0047] When the escape energy E < 1, a joint mechanism of adaptive inertial weight and random perturbation is adopted for local development optimization to enhance the fine-grained characteristic response.

[0048] Compared with the prior art, the present invention achieves the following beneficial effects:

[0049] 1. This invention proposes a unified manifold approximation and projection (UMAP) method to reduce the dimensionality of multimodal high-dimensional features. By constructing a high-dimensional adjacency graph that preserves the local structure and combining it with cross-entropy loss optimization, the geometric relationships of multi-source observation data such as remote sensing images, wind speed, and temperature and humidity features are preserved, forming a low-dimensional time series embedding with good structural consistency, providing a unified representation space for subsequent modeling.

[0050] 2. This invention proposes a recurrent neural network (VA-RNN) based on variational attention mechanism. It designs a multi-head attention module to extract key spatiotemporal feature regions and generates window aggregated features through spatial compression and aggregation. Combined with RNN for dynamic sequence learning, it realizes temporal modeling and classification prediction of rapid enhancement trends.

[0051] 3. This invention proposes an improved Harris Hawks optimization algorithm (DDHHO) for key hyperparameter search, introducing Tent chaotic initialization and nonlinear escape mechanisms to enhance the global exploration capability and local convergence speed of model structure search, and solve the problems of traditional optimization algorithms being prone to getting trapped in local optima and difficult to tune parameters.

[0052] 4. This invention proposes to simulate the nonlinear fluctuation mode during the rapid intensification process of typhoons through nonlinear escape energy design, and to realize the adaptive switching of the optimization search stage by using a dynamic energy regulation mechanism, so as to adapt to the complex nonlinear evolution of typhoons from weak to strong.

[0053] 5. In response to the interference characteristics in remote sensing images, this invention proposes a global exploration phase optimization strategy: using the Cauchy mutation mechanism to enhance the individual jumping ability and improve the robustness and generalization ability of the model to abnormal samples such as cloud image occlusion and high-latitude disturbance.

[0054] 6. This invention proposes a local development stage optimization mechanism that integrates adaptive inertial weights and random perturbations to enhance the model's response to small-scale features such as pressure gradients and sudden changes in wind speed, thereby achieving fine-grained modeling of micro-processes that can be rapidly enhanced.

[0055] In summary, this invention, through a collaborative design involving multi-source data dimensionality reduction, time-series model improvement, and optimized algorithm enhancement, can significantly improve the prediction accuracy, robustness, and practical value of rapid typhoon intensification, and has high scientific research significance and engineering application potential.

[0056] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0057] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0058] Figure 1 This is a flowchart illustrating a rapid enhancement prediction method for typhoons based on the fusion of spatiotemporal sequences and multimodal features according to an embodiment of the present invention.

[0059] Figure 2 This is a schematic diagram of the architecture of the prediction model according to an embodiment of the present invention;

[0060] Figure 3 This is a schematic diagram of the prediction model optimization process according to an embodiment of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0062] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0063] Figure 1 This is a schematic diagram of a rapid typhoon enhancement prediction method based on the fusion of spatiotemporal sequences and multimodal features according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the prediction model architecture according to an embodiment of the present invention. Figure 1 and Figure 2 As shown, a rapid typhoon enhancement prediction method 100 based on the fusion of spatiotemporal sequences and multimodal features includes the following steps:

[0064] S110: Based on the best typhoon path data and multi-source satellite observation data, construct a multimodal spatiotemporal sequence dataset and an auxiliary dataset;

[0065] The steps for constructing a multimodal spatiotemporal sequence dataset are as follows:

[0066] N consecutive typhoon observation sequences were collected at 6-hour intervals. Each sequence includes: a) Remote sensing cloud image sequence: depth feature vectors extracted from remote sensing cloud images using a convolutional neural network pre-trained on a basic cloud image dataset; b) Typhoon intensity sequence: maximum wind speed at the center; c) Upward cloud top brightness temperature sequence: infrared brightness temperature measurements; d) Temperature profile feature sequence: vertical temperature distribution features compressed by wavelet; e) Humidity profile feature sequence: vertical humidity distribution features compressed by wavelet; f) Rapid enhancement markers: binary labels indicating an intensity increase ≥ 7 m / s over the next 12 hours.

[0067] A multimodal spatiotemporal sequence dataset x represents a sequence sample (containing N consecutive time steps). This represents the i-th sample (sequence), which contains features across n time steps, and its format is: <{Remote Sensing Cloud Map 1, ..., Remote Sensing Cloud Map 1} n},{Typhoon Intensity 1, ...,Typhoon Intensity n},{Upward Cloud Top Brightness Temperature 1, ...,Upward Cloud Top Brightness Temperature n}, {Temperature profile feature 1,...,Temperature profile feature n},{Humidity Profile Feature 1, ...,Humidity Profile Feature n}, Quickly enhance the marker>. This represents the i-th typhoon observation sequence sample, which contains a feature sequence of n consecutive time steps (6-hour intervals) and a binary label indicating whether the sequence will experience rapid intensification in the next 12 hours.

[0068] Among them, the remote sensing cloud image sequence is extracted from the remote sensing cloud image by encoding a convolutional neural network (CNN) pre-trained on the basic cloud image dataset. Based on the ResNet-50 pre-trained model, its data form is the original satellite image (Himawari-8 / FY4), a 512-dimensional vector; the typhoon intensity sequence is the maximum wind speed at the center (m / s), a scalar; the uplift cloud top brightness temperature sequence is the infrared brightness temperature value (K), which is a scalar obtained after standardization; the temperature profile feature sequence is the vertical temperature distribution curve, and the feature vector obtained by Daubechies wavelet basis function compression, a 128-dimensional vector; the humidity profile feature sequence is the vertical humidity distribution curve, and the feature vector obtained by Daubechies wavelet basis function compression, a 128-dimensional vector; the rapid enhancement label is a binary label (0 / 1), for example, if the intensity increase in the next 12 hours is ≥7m / s, it is 1, otherwise it is 0.

[0069] Spatiotemporal alignment of multi-source observation data is achieved using cubic spline interpolation.

[0070] Data augmentation is performed through path simulation and feature perturbation to enhance sample diversity.

[0071] Auxiliary dataset: Typhoon life cycle identifiers (1 for rising phase, 0 for falling phase) and short-term intensity difference parameters (whether the intensity difference between time t and the previous 3 / 6 / 9 hours meets the standard) aligned with the multimodal spatiotemporal sequence data, in the format <life cycle identifier, short-term intensity difference parameter>.

[0072] S120: The Unified Manifold Approximation and Projection (UMAP) method is used to perform dimensionality reduction preprocessing on high-dimensional multimodal spatiotemporal sequence data to generate a one-dimensional time series embedding representation of typhoon observation sequences;

[0073] To fully extract potential spatiotemporal pattern information from multi-source heterogeneous features such as remote sensing cloud images, typhoon intensity sequences, uplift cloud top features, and temperature and humidity profiles, step S120 is used to perform multimodal data preprocessing. The Unified Manifold Approximation and Projection (UMAP) method is used to reduce the dimensionality of high-dimensional multimodal spatiotemporal sequence data so that subsequent time series modeling and rapid augmentation classification can be performed.

[0074] Step S121: High-dimensional space similarity modeling;

[0075] First, in a high-dimensional feature space, a weighted adjacency graph preserving local structure is constructed. Sample similarity is calculated based on weighted Euclidean distance and local neighborhood scale parameters. The local neighborhood scale parameters are negatively correlated with the rate of change of typhoon center pressure or the intensity of sudden wind speed changes, used to enhance sensitivity to rapid intensification precursors. The high-dimensional similarity between samples is defined as:

[0076] ;

[0077] in, : The multimodal observation feature vector after fusion at the i-th time step (including meteorological satellite image features, sea surface temperature, air pressure, wind speed, etc.). : The multimodal feature vector at the j-th time step, used to compare with Compare similarity. : and The weighted Euclidean distance between them measures the differences in typhoon development status at different time points in the multimodal fusion space. The local neighborhood scale parameter at the i-th time step is used to construct the scale of the neighborhood at the i-th time step. When using an adjacency graph centered on the typhoon, the scale of its local neighborhood is adjusted. This scale is typically negatively correlated with the rate of change of the central pressure or the intensity of the sudden wind speed change at time step i; the smaller the value, the better. The smaller the local neighborhood of a point, the better. The more sensitive a point is to local disturbances (such as a sharp drop in central air pressure or a sudden increase in wind speed). : and The similarity score reflects whether the typhoon situations at two time steps are similar or continuous. This score serves as the edge weight for constructing a high-dimensional adjacency graph.

[0078] This similarity function, by introducing a local Gaussian kernel mechanism, enables the model to perceive the nonlinear local perturbation structure during the rapid intensification of a typhoon, inspired by the concept of hyperbolic space geometry.

[0079] Step S122: Low-dimensional space mapping modeling;

[0080] In a low-dimensional embedding space, the similarity between point pairs is defined. To preserve the structural information of the original space:

[0081]

[0082] in, The distance is Euclidean distance, used to measure the evolution trajectory of a typhoon's development in a low-dimensional space. : The low-dimensional embedding representation of the i-th time step, which is a one-dimensional scalar representing the embedding characteristics of the typhoon's overall situation at that time point. : The low-dimensional embedding representation at time step j, which participates in similarity modeling as a neighbor of i. The normalized low-dimensional similarity score reflects whether the embedding space retains the proximity relationships from the original high-dimensional space. Its form is similar to the softmax distribution and is used to optimize structure preservation.

[0083] This low-dimensional similarity function is resistant to outliers and gradient explosion, and is suitable for non-Euclidean temporal embedding modeling. It helps to preserve the geometric neighborhood information of the "typhoon intensity inflection point" and "rapid intensification situation", which can enhance the model's ability to represent the rapid intensification phase of extreme typhoons.

[0084] Step S123: Similarity Preservation Optimization;

[0085] The embedding representation is optimized by minimizing the cross-entropy loss due to the difference in similarity between high- and low-dimensional spaces. The low-dimensional embedding is then iteratively optimized to generate a one-dimensional time series embedding representation, ensuring that the structure of the original multimodal information remains as consistent as possible after dimensionality reduction.

[0086] ;

[0087] in, The overall loss function measures the consistency of the dimensionality-reduced embedded sequences in preserving the original multimodal observation structure of the typhoon. A smaller loss indicates that the low-dimensional sequence more accurately represents the similarity relationships of the high-dimensional features. The similarity between the i-th and j-th time points in the high-dimensional multimodal feature space (see the previous formula) reflects the degree of closeness of the typhoon development status in the original feature space. The similarity between the i-th and j-th time points in the low-dimensional embedding space (see the previous formula) reflects whether the time embedding after dimensionality reduction maintains the original state relationship. : Represents the penalty term for the original similarity information on the low-dimensional similarity bias, used to strengthen structure preservation. It is used to suppress the erroneous aggregation of time points that are originally "far apart" by low-dimensional embeddings, and improve the distinguishability between dissimilar points. : Different observation time points in the time series (such as the typhoon center location, intensity level, wind field distribution, etc. at different times) satisfy This is used to avoid double counting.

[0088] The optimization process uses the Adam optimizer iteratively to obtain a typhoon development time series embedding with low dimensionality but good structural consistency, i.e., a one-dimensional time series embedding representation. :

[0089]

[0090] in This is the embedding representation of the typhoon observation sequence that integrates multimodal features at time step t. By using the UMAP dimensionality reduction method described above, the typhoon observation sequence that integrates multimodal features is converted into a one-dimensional time series embedding representation, providing an input basis for subsequent time series-based modeling and rapid enhancement prediction.

[0091] S130: The one-dimensional time series embedding representation and auxiliary data are used as independent input channels and input into the trained typhoon observation network model to predict the probability of rapid typhoon intensification.

[0092] The typhoon observation network model is a multimodal spatiotemporal fusion deep learning architecture. Its core consists of a variational attention recurrent neural network (VA-RNN), and its hyperparameters are optimized using an improved Harris Eagle Optimization (DDHHO) algorithm.

[0093] Step S131: The one-dimensional time series embedding representation Z and the auxiliary data A are used as independent input channels and are used as inputs to the VA-RNN.

[0094] For each time step t, the auxiliary data input vector : [Lifecycle Identifier, Short-term Intensity Difference Parameter] = [Lifecycle Identifier_t, ΔIntensity3h_t, ΔIntensity6h_t, ΔIntensity9h_t], where, Lifecycle Identifier_t: the typhoon lifecycle identifier at time t, 1 indicates the rising phase, 0 indicates the falling phase; ΔIntensity3h_t, ΔIntensity6h_t, ΔIntensity9h_t: the intensity difference between time t and the previous 3, 6, and 9 hours (or a binary variable indicating whether the standard is met). Optionally, the Short-term Intensity Difference Parameter is three binary variables (0 / 1) indicating whether the standard is met.

[0095] The variational attention recurrent neural network has independent input channels: a) the first channel inputs the dimensionality-reduced temporal embedding sequence; b) the second channel inputs auxiliary data;

[0096] Dynamically integrate dual-channel outputs through a gating fusion mechanism:

[0097] ;

[0098] in, These represent the outputs of the first type of attention head, the first type of attention head, and the first type of attention head, respectively. This is the internal weighting coefficient (specifically, it can be 0.6). This represents the one-dimensional time series embedding feature sequence within a sliding window of length L (Window_Length) centered at time step t (or ending at that time step). The feature sequence has the following shape: ), It is its characteristic dimension; To indicate and A sequence of auxiliary data feature windows (also of the same shape) that are strictly time-aligned and have the same window length L. ), It is its characteristic dimension; This is a dynamic gating coefficient, whose value is adaptively adjusted based on the status of the auxiliary data.

[0099] When auxiliary data detects an intensity transition At that time, the gating coefficient This enhances the effect of cross-attention;

[0100] When the auxiliary data shows a stable state This enhances the dominance of temporal embedding.

[0101] In embodiments of the present invention, the auxiliary data includes: a typhoon lifecycle identifier, encoded as a binary scalar; and multi-scale intensity difference parameters, encoded as unique heat vectors. Specifically:

[0102] Auxiliary data feature vector Time step t includes: typhoon lifecycle identifier (1-dimensional scalar, 1 = rising phase, 0 = falling phase), and Δintensity 3h_t, Δintensity 6h_t, and Δintensity 9h_t. Therefore, For L consecutive elements within the window The sequence is composed of shapes [L, 4].

[0103] Step S132: Typhoon observation network model training;

[0104] The training process of the typhoon observation network model employs a variational attention-based recurrent neural network (VA-RNN) to process one-dimensional time-series embeddings and auxiliary data, including:

[0105] Step S1321: Input the one-dimensional time series embedding representation and auxiliary data into the multi-head attention mechanism, and generate query vector, key vector and value vector through linear transformation to capture the correlation of key regions in different time steps;

[0106] First, the input one-dimensional time series embedding representation Z and auxiliary data A are mapped into query (Q), key (K), and value (V) vectors to capture the correlation of key regions in different time steps:

[0107] ;

[0108] : These are query, key, and value vector matrices, respectively, representing linear transformations of features from different time points or modalities; : The dimension of the key vector, used for scaling to ensure numerical stability.

[0109] This attention mechanism can automatically identify important spatiotemporal feature patterns that are associated with rapid future enhancement.

[0110] Step S1322: Employ a parallel multi-head structure to fuse multi-scale features: Each attention head captures signals from different physical mechanisms, including features of strong convective cloud regions or abrupt changes in temperature and humidity profiles; The multi-head outputs are spliced ​​together and fused across heads using a linear mapping matrix to obtain a fused feature map;

[0111] Multi-head attention mechanisms include three specific types of attention heads, each consisting of multiple parallel heads, which, when generating query (Q), key (K), and value (V) vectors, include:

[0112] (1) Multi-head attention mechanisms include three types of dedicated attention heads:

[0113] Type 1 attention heads (temporal embedding heads): Mz in total, only accepting one-dimensional time series embedding feature window sequences. Its query (Q), key (K), and value (V) vectors are weighted by a dedicated weight matrix. (Initialized according to 𝒩(0,0.1)) Projected independently. Feature vectors are generated at each time step to capture abrupt changes in the typhoon's development process;

[0114] Type II attention header (auxiliary data header): Ma headers in total, only receiving auxiliary data feature window sequences. The query (Q), key (K), and value (V) vectors are used via a dedicated weight matrix. (Initialized to follow 𝒩(0,0.05)) Projected independently. Feature vectors are generated at each time step to monitor intensity transition signals;

[0115] The third type of attention head (cross-modal head): a total of Mc heads, which simultaneously receive one-dimensional time series embedded feature window sequences. and auxiliary data feature window sequence Its query (Q) vector is passed through the weight matrix. Projecting independently Feature vectors are generated at each time step; their key (K) and value (V) vectors are weighted by a weight matrix. Projecting independently The feature vector is generated at each time step to capture the cross-pattern of the two types of features;

[0116] The multi-head mechanism helps to comprehensively capture key enhancement signals at different scales and regions. By using three types of attention heads, the model adaptively determines which time steps rely more on temporal embedded features (e.g., when the typhoon is in a stable phase) and which time steps rely more on auxiliary data (e.g., when an intensity transition is detected). This retains the multi-head attention mechanism of the original scheme while enhancing the model's ability to fuse multi-source data.

[0117] Concatenate the outputs head_m (m=1, ..., M) of all the attention heads mentioned above, where M = Mz + Ma + Mc.

[0118] ;

[0119] in, : The output of the m-th attention head, focusing on a specific time window, spatial region, or physical variable (such as a strong convective cloud region or a wet stratum structure); M: The number of attention heads; : Linear mapping matrix, used to uniformly map the concatenated multi-head results to the dimensions required by the model; Vector concatenation operation; This represents the fused feature map output by the multi-attention mechanism.

[0120] Step S1323: Perform temporal window compression aggregation on the fused feature map, and generate window aggregated features by global average pooling along the time dimension. ;

[0121] set up Let be the multi-head attention output tensor corresponding to time step t, and its shape is: Batch represents the number of samples in a batch, and L represents the length of the input window (number of time steps). Let the feature dimension of the attention output be , and let the output dimension of each attention head be . The concatenated feature dimensions For each sample's window sequence, perform global average pooling along the time dimension to generate window aggregated features. This is used for subsequent sequence modeling in RNNs.

[0122] ;

[0123] in, It represents the aggregated feature of the entire window (containing L time steps) ending at time step t, which encodes the comprehensive evolution state of the typhoon within the window [t-L+1, t]. , represents the feature vector of the multi-head attention output tensor at the τ-th position in the window sequence corresponding to time step t (i.e., For time steps t < L, a complete window is constructed using zero-padding.

[0124] Step S1324: Input the window aggregation features into a recurrent neural network to model the temporal evolution, and encode the typhoon energy accumulation process through hidden states; input the hidden state of the last time step into the classifier to output the rapid enhancement probability for the next 12 hours.

[0125] The window-aggregated feature sequence actually contains features from time step L to T. (Where T=N, i.e., the same length as the original observation sequence) The data is sequentially input into a recurrent neural network (RNN) to model the temporal evolution and capture the dynamic changes over time.

[0126] ;

[0127] : Multimodal embedding feature representation of the typhoon at time t (such as fused wind speed field, sea surface temperature, air pressure structure and other information). : The hidden state vector at time t, which encodes the historical evolution information and current state of the typhoon at the current time point. : The hidden state at time t-1, representing the memory of the previous typhoon evolution trend. : Recurrent neural network units used to model the time-dependent features of typhoon development. During training, the hidden state at each time step needs to be saved. .

[0128] Finally, the state will be hidden at the last moment. Input to a classifier (fully connected layer + Sigmoid) and output fast augmentation probability. :

[0129] ;

[0130] in, The global time series representation of the typhoon at the last moment (i.e. the moment before the predicted target), which incorporates all previous observation states. : The weight matrix of the classifier, used to linearly map the hidden states to the prediction space. : The bias term of the classifier. The Sigmoid function is used to compress the linear output to the [0,1] interval, representing the prediction probability of a rapid increase in the typhoon sample. : The confidence level of rapid enhancement predicted by the model. The closer the value is to 1, the higher the probability of rapid enhancement.

[0131] The classifier consists of a fully connected layer and a sigmoid activation function, outputting probability values; and the probability threshold is set to 0.55. If so, the typhoon sample is determined to be "rapidly intensifying".

[0132] The total loss function for model training consists of a binary cross-entropy loss function and a loss function that includes feature reconstruction constraints. The total loss function is:

[0133] ;

[0134] in: To reconstruct the loss weighting coefficients, dynamic adjustments can be made based on the typhoon type, with values ​​ranging from 0.05 to 0.2. When a typhoon eyewall remodeling event is detected, The value can be temporarily increased to 0.25 to strengthen the feature preservation constraint. Reconstruct the loss function. :

[0135] ;

[0136] To reconstruct the mapping function and satisfy: , For dimension , To hide the state dimension, Embedding dimensions for one-dimensional time series; For the fully connected layer mapping function, the hidden state is... Mapped to a one-dimensional time series embedding representation at time step t Same dimensions; The layer normalization function is used; the reconstruction target is the embedding at the end time t of the window. .

[0137] Binary cross-entropy loss for:

[0138] ;

[0139] in, : Real and fast tag enhancement; Model prediction confidence.

[0140] Step S133: Optimize the typhoon observation network model;

[0141] To improve the generalization performance and accuracy of spatiotemporal sequence and multimodal feature fusion networks in the task of rapid typhoon enhancement prediction, this method introduces an improved Harris Hawks optimization algorithm (DDHHO) to automatically search for key hyperparameters of the model (such as the number of recurrent layer units, learning rate, number of multi-head attention heads, feature fusion weights, etc.), aiming to solve the problems of slow convergence speed and easy getting trapped in local optima in traditional methods.

[0142] Specifically, the typhoon observation network model optimization process involves hyperparameter optimization using the improved Harris Hawks optimization algorithm (DDHHO), including:

[0143] Step S1331: Initialize the population using the Tent chaotic map combined with random perturbation;

[0144] In the traditional HHO algorithm, the initialization of the optimized population uses a random uniform distribution, which can easily lead to individuals clustering in local regions of the typhoon sample feature space, affecting the global modeling of the coupling relationships of complex multimodal features (such as cloud texture, sea surface temperature distribution, and air pressure changes). DDHHO uses a Tent chaotic mapping combined with random factors to generate the initial population, enhancing distribution diversity.

[0145] ; The position of the nth individual in the Tent map represents the chaotic state of the individual in the normalized feature space. The position of the next generation of individuals in the Tent map has better divergence and traversal. :exist Random perturbation tops sampled within the interval increase the uncertainty and diversity of the initial population; N: population size in the optimization algorithm, representing the number of candidate individuals participating in the hyperparameter search of the typhoon prediction model.

[0146] This formula is used to generate a chaotic sequence of initial optimized individuals. To enhance the distribution diversity of the population in a complex multimodal feature space, the Tent mapping and random perturbation are introduced to enable the optimized population to more fully cover the complex combination space of multimodal features such as cloud texture, sea surface temperature, and air pressure, thus avoiding getting trapped in local optima.

[0147] After generating chaotic individuals, the specific hyperparameter values ​​are obtained through inverse mapping:

[0148] ;

[0149] The model hyperparameter values ​​corresponding to the nth individual, such as the dimension of the recurrent layer, the number of attention heads, the size of the convolutional kernel, etc. The minimum range of values ​​for this hyperparameter represents the starting point under the simplified structure of the model; The maximum range of this hyperparameter represents the upper bound of the model's complex structure; The normalized individual location value after Tent mapping controls the mapping weight.

[0150] This formula maps the results of chaos. Mapped to specific model hyperparameter values This is used to control the network structure and training configuration. Through... The mapping can generate various network structure configurations for training the "Typhoon Rapid Enhancement Prediction Model" and iteratively search for the optimal structure in the DDHHO framework to improve the model's perception and generalization ability of modal coupling changes during typhoon evolution.

[0151] Step S1332: Design a nonlinear escape energy function to dynamically regulate the ratio of global exploration to local exploitation;

[0152] Figure 3 This is a schematic diagram of the prediction model optimization process according to an embodiment of the present invention. Figure 3 As shown, DDHHO performed a nonlinear reconstruction of the escape energy E to simulate the characteristic of uncertainty gradually decreasing over time during typhoon development. The escape energy E controls the ratio of global to local search, and its update formula is as follows:

[0153] r: Current iteration number, representing the time progress of the algorithm evolution; : Maximum number of iterations, representing the time endpoint of the optimization process; : Nonlinear mapping adjustment factor, used to control the slope and directionality of the escape energy curve, reflecting the phased strengthening trend of the typhoon evolution process; E: Escape energy, which determines the weight ratio of global exploration and local convergence in the DDHHO search process, used to simulate the nonlinear uncertainties (such as abrupt development or delayed energy release) in the typhoon strengthening process.

[0154] This nonlinear mapping aims to achieve a smooth transition from early high exploratory (E≈1, emphasizing global modeling of large-scale relationships between multimodal information) to later high exploitation (E≈0, improving local convergence ability and integrating high-confidence feature representations), effectively adapting to the shift in decision-making focus from breadth search to depth mining during the rapid enhancement and evolution of typhoons.

[0155] Step S1333: Search strategy optimization;

[0156] (1) Optimization of the global exploration phase

[0157] When the escape energy E≥1, the Cauchy mutation mechanism is used to update the parameters for global exploration and optimization, thereby enhancing the ability to escape local optima.

[0158] To address the significant noise and inconsistencies in the multimodal feature space (such as cloud cover interference in remote sensing images), DDHHO introduces a Cauchy mutation mechanism to perturb individual eagles, increasing their probability of escaping local optima.

[0159] ;

[0160] The parameter vector representing the current individual indicates the parameter position of a certain optimization candidate solution in the model structure or fusion strategy (such as attention fusion weights, sequence modeling hyperparameters). : Perturbation intensity coefficient, used to adjust the variation amplitude to adapt to the distribution density of different modal features (e.g., areas with drastic changes in image texture require larger perturbations). Among them, the cloud image interference index is calculated based on the proportion of cloud-covered area in the satellite image; Standard Cauchy distribution samples have thick tails, allowing for larger step sizes of jump perturbations, thus widening the search radius and helping to overcome feature traps caused by cloud cover, local anomalies, etc. in remote sensing images.

[0161] This strategy enhances the global exploration jump capability, avoiding getting trapped in local optima in complex multimodal feature spaces, and is particularly suitable for modeling typhoon structural abrupt changes or extreme enhancement samples. This mechanism combines chaotic initialization and nonlinear escape energy design to strengthen the model's ability to identify feature abrupt changes during the sudden enhancement phase of typhoons, improve robustness to scenarios such as remote sensing interference and high-latitude disturbances, and enhance the fitting accuracy of feature fusion representation.

[0162] (2) Optimization of the local development stage

[0163] When the escape energy E < 1, a joint mechanism of adaptive inertial weight and random perturbation is adopted for local development optimization to enhance the fine-grained characteristic response.

[0164] To enhance local detail modeling (such as the response of pressure gradients and wind speed differences at small scales), a joint mechanism of adaptive inertial weights and random perturbations is introduced. The inertial weight adjustment formula is as follows:

[0165] ;

[0166] Inertia weight is used to control the update magnitude and inertia trend of the current individual parameters, affecting the fitting accuracy of small-scale meteorological features (such as local air pressure changes and wind speed differences) during the optimization process. : Upper limit of inertial weight, usually set to a large value, is used to optimize the initial rapid convergence to the potential region and ensure that the model can cover the global feature distribution in the typhoon multimodal data; : Lower bound of inertia weight, set to a small value, used to optimize the local sensitivity to key enhancement signals in the later stage, and refine microstructures such as "typhoon eyewall remodeling" or "deep convection burst"; r: current iteration number, representing the time node in the model learning process; The maximum number of iterations determines the total optimization step size and the inertia decay rate.

[0167] The upper and lower limits of the adaptive inertia weight are related to the typhoon type: Oceanic typhoons: Landfall-type typhoons: It is used to adapt to the sensitivity of microscopic features of different typhoon intensification processes.

[0168] This formula is used to dynamically adjust the inertia strength of search individuals during the update process, so as to achieve a smooth transition from global exploration to local development. It is particularly suitable for capturing micro-disturbance signals during the intensification process of typhoons. Large inertia weights ensure that the model has stronger exploratory power in the initial stage and improves the convergence speed; while small inertia weights are suitable for focusing on strengthening detailed modeling and boundary sample discrimination in the later stage, such as short-term wind pressure anomalies before the sudden intensification of a typhoon.

[0169] Random perturbations during the local development phase are distributed using a Gaussian distribution:

[0170] ;

[0171] in, It is the globally optimal solution found up to the r-th iteration. It is the solution vector of the r-th iteration. It is a random noise vector. The data follows a multivariate normal distribution with a mean of 0 and a covariance matrix of . ; The identity matrix is ​​used to ensure that each dimension of the noise is independent; the standard deviation σ is negatively correlated with the weak signal strength of the typhoon enhancement precursor, and the specific value is σ=0.1×(1-the recall rate of the current best model on the validation set).

[0172] During the local search phase, a joint random perturbation mechanism is introduced to apply slight perturbations to the current individual model. This aims to escape local minima and enhance the model's responsiveness to real rapid enhancement signals. It controls whether to introduce perturbation jumps when an individual model converges to a local optimum, reducing the risk of falling into "false enhancements" or "misjudging weak signals" and improving the model's robustness in identifying real rapid enhancement behavior. The local perturbation mechanism is particularly suitable for the fuzzy prediction phase before typhoon intensification. By probabilistically perturbing the model, it stimulates the search activity in boundary regions, thereby identifying potential signs of explosive intensification (such as anomalous cloud clustering near high sea surface temperatures).

[0173] The DDHHO optimization strategy comprehensively considers the nonlinear, multi-scale, and multi-modal evolution characteristics of rapid typhoon intensification, effectively improving the quality of model hyperparameter configuration, and thus significantly enhancing the model's ability to fit complex meteorological processes and its accuracy in discrimination.

[0174] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0175] It should also be noted that, in the embodiments of this application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0176] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in the embodiments of this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown in this application, but is to be accorded the widest scope consistent with the principles and novel features disclosed in the embodiments of this application.

Claims

1. A rapid enhancement prediction method for typhoons based on the fusion of spatiotemporal sequences and multimodal features, characterized in that, Includes the following steps: Based on typhoon optimal path data and multi-source satellite observation data, a multimodal spatiotemporal sequence dataset and an auxiliary dataset are constructed. A unified manifold approximation and projection method is used to perform dimensionality reduction preprocessing on high-dimensional multimodal spatiotemporal sequence data, generating a one-dimensional time series embedding representation of typhoon observation sequences; The one-dimensional time series embedding representation and auxiliary data are used as independent input channels and input into the trained typhoon observation network model to predict the probability of rapid typhoon intensification. The typhoon observation network model is a multimodal spatiotemporal fusion deep learning architecture, whose core consists of a variational attention recurrent neural network and is optimized for hyperparameters using an improved Harris Hawks optimization algorithm. The training process of the typhoon observation network model employs a recurrent neural network with a variational attention mechanism to process the one-dimensional time series embedding representation and the auxiliary data, specifically including: The one-dimensional time series embedding representation and the auxiliary data input multi-head attention mechanism are used to generate query vectors, key vectors and value vectors through linear transformation, which are used to capture the correlation of key regions in different time steps; A parallel multi-head structure is adopted to fuse multi-scale features: each attention head captures signals from different physical mechanisms, including features of strong convective cloud regions or abrupt changes in temperature and humidity profiles; the outputs of multiple heads are spliced ​​and fused across heads through a linear mapping matrix to obtain a fused feature map; Perform temporal window compression aggregation on the fused feature map, and generate window aggregated features by global average pooling along the time dimension; The window aggregation features are input into a recurrent neural network to model the temporal evolution, and the typhoon energy accumulation process is encoded through hidden states; the hidden state of the last time step is input into a classifier to output the rapid enhancement probability for the next 12 hours; The multi-head attention mechanism of the variational attention recurrent neural network includes three types of dedicated attention heads: Type I attention heads: Mz in total, only accepting one-dimensional time-series embedded feature window sequences. It is used to capture abrupt changes during the development of typhoons; Type II attention heads: Ma heads in total, only receiving auxiliary data feature window sequences. Used to monitor intensity transition signals; The third type of attention head: a total of Mc heads, simultaneously receiving a one-dimensional time series embedded feature window sequence. and auxiliary data feature window sequence This is used to capture the cross-pattern of two types of features; The variational attention recurrent neural network dynamically integrates the dual-channel outputs through a gating fusion mechanism. in, These represent the outputs of the first type of attention head, the second type of attention head, and the third type of attention head, respectively. These are internal weighting coefficients; This is a dynamic gating coefficient, whose value is adaptively adjusted according to the state of the auxiliary data: when the auxiliary data detects an intensity transition... At that time, the gating coefficient When auxiliary data shows a stable state, ; Among these, hyperparameter optimization is performed using an improved Harris Hawks optimization algorithm, including: The population is initialized using Tent chaotic mapping combined with random perturbation; Design a nonlinear escape energy function to dynamically regulate the ratio of global exploration to local exploitation; When the escape energy E≥1, the Cauchy mutation mechanism is used to update the parameters for global exploration and optimization, thereby enhancing the ability to escape local optima. When the escape energy E < 1, a joint mechanism of adaptive inertial weight and random perturbation is adopted for local development optimization to enhance the fine-grained characteristic response.

2. The typhoon rapid enhancement prediction method based on spatiotemporal sequence and multimodal feature fusion according to claim 1, characterized in that, The steps for constructing the multimodal spatiotemporal sequence dataset are as follows: N consecutive typhoon observation sequences were collected at 6-hour intervals. Each sequence included: a) Remote sensing cloud image sequence: depth feature vectors extracted from remote sensing cloud images using a convolutional neural network pre-trained on a basic cloud image dataset; b) Typhoon intensity sequence: maximum wind speed at the center; c) Upward cloud top brightness temperature sequence: infrared brightness temperature measurements; d) Temperature profile feature sequence: vertical temperature distribution features compressed by wavelet; e) Humidity profile feature sequence: vertical humidity distribution features compressed by wavelet; f) Rapid enhancement markers: binary labels indicating an intensity increase ≥ 7 m / s over the next 12 hours. Spatiotemporal alignment of multi-source observation data is achieved using cubic spline interpolation. Data augmentation is achieved through path simulation and feature perturbation.

3. The typhoon rapid enhancement prediction method based on spatiotemporal sequence and multimodal feature fusion according to claim 2, characterized in that, The specific process of the dimensionality reduction preprocessing includes: A weighted adjacency graph that preserves local structure is constructed in a high-dimensional space, and sample similarity is calculated based on weighted Euclidean distance and local neighborhood scale parameters. The local neighborhood scale parameters are negatively correlated with the rate of change of air pressure at the typhoon center or the intensity of sudden wind speed changes, which is used to enhance the sensitivity to rapid intensification precursors. Define a similarity function for robust anomalies in a low-dimensional space; The embedding representation is optimized by minimizing the cross-entropy loss of the similarity difference between high- and low-dimensional spaces, and a one-dimensional time series embedding representation is generated through iterative optimization.

4. The typhoon rapid enhancement prediction method based on spatiotemporal sequence and multimodal feature fusion according to claim 3, characterized in that, The classifier consists of a fully connected layer and a sigmoid activation function, and outputs a probability value; the probability judgment threshold is set to 0.55, and when the probability value exceeds 0.55, it is judged as fast augmentation; The total loss function for model training consists of a binary classification cross-entropy loss function and a temporal embedding reconstruction loss function.

5. The typhoon rapid enhancement prediction method based on spatiotemporal sequence and multimodal feature fusion according to claim 4, characterized in that, The total loss function for training the model is: in: The weighting coefficients for the temporal embedding reconstruction loss range from 0.05 to 0.2; the temporal embedding reconstruction loss function... : To reconstruct the mapping function and satisfy: , For dimension The weight matrix, To hide the state dimension, Embedding dimensions for one-dimensional time series; For the fully connected layer mapping function, the hidden state is... Mapped to a one-dimensional time series embedding representation at time step t Same dimensions; For layer normalization function; Binary cross-entropy loss for: in, : Real and fast tag enhancement; Model prediction confidence.

6. The typhoon rapid enhancement prediction method based on spatiotemporal sequence and multimodal feature fusion according to claim 5, characterized in that, in, The auxiliary data includes: typhoon life cycle identifiers, encoded as binary scalars; and multi-scale intensity difference parameters, encoded as unique heat vectors.

Images