Typhoon rapid enhancement prediction method based on time-space sequence and multi-modal feature fusion

Abstract

The invention relates to the technical field of typhoon prediction, and discloses a typhoon rapid enhancement prediction method based on time-space sequence and multi-modal feature fusion, and the method comprises the steps: constructing a multi-modal time-space sequence data set and an auxiliary data set based on typhoon optimal path data and multi-source satellite observation data; a unified manifold approximation and projection method is adopted to carry out dimension reduction preprocessing on the high-dimensional multi-modal space-time sequence data, and one-dimensional time sequence embedding representation of the typhoon observation sequence is generated; taking the one-dimensional time sequence embedded representation and the auxiliary data as independent input channels, and inputting a trained typhoon observation network model to predict a typhoon rapid enhancement probability; wherein the typhoon observation network model is a multi-mode time-space fusion deep learning architecture, the core of the typhoon observation network model is composed of a variational attention recurrent neural network, and hyper-parameter optimization is carried out through an improved Harris eagle optimization algorithm. According to the invention, accurate and robust identification of the typhoon rapid enhancement process is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of typhoon prediction, and in particular to a typhoon rapid intensification prediction method based on the fusion of spatiotemporal sequences and multimodal features. Background Art

[0002] Typhoons are one of the world's most destructive extreme weather disasters. Their rapid intensification is often accompanied by a sudden drop in central pressure and a sharp increase in wind speed, posing a significant threat to maritime shipping, coastal urban safety, and emergency response deployments. Because typhoon formation and evolution are highly complex and influenced by multiple sources of observational data (such as remote sensing cloud images, meteorological satellites, sea temperature, pressure fields, and temperature and humidity profiles), accurately predicting whether a typhoon will rapidly intensify in the future is extremely technically challenging.

[0003] Current mainstream methods are mostly based on statistical features or single model structures, and have the following shortcomings: 1. Difficulty in fusing multimodal observation data: remote sensing images, meteorological parameters, cloud top brightness temperature and other feature dimensions and scales vary, making it difficult to uniformly model them; 2. Limited expression of spatiotemporal evolution characteristics: traditional methods are insufficient in dealing with the nonlinear dynamics of typhoon evolution over time, especially the inability to accurately identify key turning points before intensification; 3. Weak model generalization and optimization capabilities: conventional neural network structures have problems such as unstable training, strong hyperparameter dependence, and easy falling into local optimality.

[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 to achieve accurate and robust identification of the rapid intensification process of typhoons. Summary of the Invention

[0005] The present invention aims to address the deficiencies of existing typhoon rapid intensification prediction methods in terms of multimodal data fusion, feature expression and optimization capabilities, and proposes a typhoon rapid intensification prediction method based on the fusion of spatiotemporal series and multimodal features.

[0006] To achieve the above objectives, the following technical solutions are adopted:

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

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

[0009] The unified manifold approximation and projection (UMAP) method is used to preprocess the high-dimensional multimodal spatiotemporal series data to reduce the dimensionality and generate a one-dimensional time series embedding representation of the typhoon observation series.

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

[0011] Among them, the typhoon observation network model is a multimodal spatiotemporal fusion deep learning architecture, whose core is composed of a variational attention recurrent neural network (VA-RNN), and hyperparameter optimization is performed through the improved Harris Hawks optimization algorithm (DDHHO).

[0012] Furthermore, the steps for constructing the multimodal spatiotemporal sequence dataset are as follows:

[0013] N consecutive typhoon observation sequences are collected at 6-hour intervals. Each sequence contains: a) remote sensing cloud image sequence: deep feature vectors extracted from remote sensing cloud images using a convolutional neural network pre-trained with a basic cloud image dataset; b) typhoon intensity sequence: maximum central wind speed values; c) upthrust cloud top brightness temperature sequence: infrared brightness temperature measurements; d) temperature profile feature sequence: vertical temperature distribution features compressed by wavelet analysis; e) humidity profile feature sequence: vertical humidity distribution features compressed by wavelet analysis; f) rapid enhancement label: a binary label indicating an intensity increase of ≥7 m / s over the next 12 hours.

[0014] Cubic spline interpolation is used to achieve spatiotemporal alignment of multi-source observation data;

[0015] Data enhancement is performed through path simulation and feature perturbation to improve 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 high-dimensional space, and sample similarity is calculated based on weighted Euclidean distance and a local neighborhood scale parameter; wherein the local neighborhood scale parameter is negatively correlated with the rate of change of typhoon central pressure or the intensity of wind speed mutation, and is used to enhance sensitivity to precursors of rapid intensification;

[0018] Define a similarity function against outliers in 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 the one-dimensional time series embedding representation is generated by iterative optimization.

[0020] Furthermore, the training process of the typhoon observation network model uses 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] Inputting the one-dimensional time series embedding representation and the auxiliary data into a multi-head attention mechanism, generating a query vector, a key vector, and a value vector through linear transformation to capture the correlation of key regions in different time steps;

[0022] A parallel multi-head architecture is used to fuse multi-scale features: each attention head captures signals from different physical mechanisms, such as strong convective cloud characteristics or sudden changes in temperature and humidity profiles; multiple head outputs are spliced ​​and fused across heads using a linear mapping matrix to produce a fused feature map.

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

[0024] The window aggregation features are input into a recurrent neural network to model 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 probability of rapid enhancement in the next 12 hours.

[0025] Furthermore, the classifier is composed of a fully connected layer and a Sigmoid activation function, and outputs a probability value; and the probability judgment threshold is set to 0.55. When the probability value exceeds 0.55, it is judged as rapid enhancement;

[0026] The total loss function of model training consists of the binary cross entropy loss function and the time series 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] The first type of attention head: a total of Mz, only receives one-dimensional time series embedding feature window sequence , used to capture the sudden change characteristics during the development of typhoons;

[0029] The second type of attention head: a total of Ma heads, only receiving auxiliary data feature window sequence , used to monitor the intensity transition signal;

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

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

[0032]

[0033] in: is the time series embedding reconstruction loss weight coefficient, ranging from 0.05 to 0.2; the time series embedding reconstruction loss function :

[0034]

[0035] is the reconstructed mapping function and , The dimension is , is the hidden state dimension, Embed dimension for one-dimensional time series; is the mapping function of the fully connected layer, which transforms the hidden state Mapped to a one-dimensional time series embedding representation of the t-th time step Same dimensions; is the layer normalization function;

[0036] Binary cross entropy loss for:

[0037] ;

[0038] in, : Real fast enhanced labeling; : Model prediction confidence.

[0039] Furthermore, the auxiliary data includes: a typhoon life cycle identifier, encoded as a binary scalar; and a multi-scale intensity difference parameter, encoded as a one-hot vector.

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

[0041] ;

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

[0043] Furthermore, the improved Harris Hawks optimization algorithm is used to optimize hyperparameters, including:

[0044] Use Tent chaotic map combined with random perturbations to initialize the population;

[0045] Design a nonlinear escape energy function to dynamically control 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 optimization to enhance the ability to escape from the local optimum;

[0047] When the escape energy E<1, a combined mechanism of adaptive inertia weight and random perturbation is used to perform local development optimization and enhance fine-grained feature response.

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

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

[0050] 2. This paper proposes a recurrent neural network (VA-RNN) based on the variational attention mechanism, designs a multi-head attention module to extract key spatiotemporal feature areas, and generates window aggregation features through spatial compression aggregation. It combines RNN for dynamic sequence learning to achieve time series modeling and classification prediction of rapid enhancement trends.

[0051] 3. This paper proposes an improved Harris Hawks optimization algorithm (DDHHO) for key hyperparameter search, introduces Tent chaos initialization and nonlinear escape mechanism, improves the global exploration capability and local convergence speed of model structure search, and solves the problem that traditional optimization algorithms are prone to falling into local optimality and difficult to adjust parameters.

[0052] 4. The present invention proposes to simulate the nonlinear fluctuation pattern during the rapid intensification of typhoons through nonlinear escape energy design, and use dynamic energy control mechanism to achieve adaptive switching in the optimization search stage to adapt to the complex nonlinear evolution of the typhoon intensification process from weak to strong.

[0053] 5. Aiming at the interference characteristics in remote sensing images, the present invention proposes a global exploration phase optimization strategy: the Cauchy mutation mechanism is used to enhance the individual jumping ability, thereby improving the model's robustness and generalization ability to abnormal samples such as cloud occlusion and high-latitude disturbances.

[0054] 6. The present invention proposes a local development stage optimization mechanism that integrates adaptive inertia weights and random perturbations to enhance the model's ability to respond to small-scale features such as pressure gradients and wind speed mutations, thereby achieving rapid enhancement of fine-grained modeling of microscopic processes.

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

[0056] It should be understood that the contents described in the summary of the invention are not intended to limit the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The above and other features, advantages and aspects of the embodiments of the present invention will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. The accompanying drawings are provided for a better understanding of the present invention and do not constitute a limitation of the present invention. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, among which:

[0058] Figure 1 This is a flow chart of a method for predicting rapid typhoon intensification based on the fusion of spatiotemporal sequences and multimodal features according to an embodiment of the present invention;

[0059] Figure 2 Schematic diagram of the prediction model architecture of an embodiment of the present invention;

[0060] Figure 3 2 is a schematic diagram of the prediction model optimization process of an embodiment of the present invention. DETAILED DESCRIPTION

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0062] In this document, the term "and / or" simply describes a relationship between related objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document generally indicates that the related objects are in an "or" relationship.

[0063] Figure 1 This is a flow chart of a method for predicting rapid typhoon intensification based on the fusion of spatiotemporal sequences and multimodal features according to an embodiment of the present invention; Figure 2 Schematic diagram of the prediction model of the embodiment of the present invention. Figure 1 and Figure 2 As shown, a typhoon rapid intensification prediction method 100 based on the fusion of spatiotemporal sequence and multimodal features includes the following steps:

[0064] S110: Construct multimodal spatiotemporal series datasets and auxiliary datasets based on typhoon optimal path data and multi-source satellite observation data;

[0065] Among them, the steps for constructing the multimodal spatiotemporal sequence dataset are:

[0066] N consecutive typhoon observation sequences are collected at 6-hour intervals. Each sequence contains: a) remote sensing cloud image sequence: deep feature vectors extracted from remote sensing cloud images using a convolutional neural network pre-trained with a basic cloud image dataset; b) typhoon intensity sequence: maximum central wind speed values; c) upthrust cloud top brightness temperature sequence: infrared brightness temperature measurements; d) temperature profile feature sequence: vertical temperature distribution features compressed by wavelet analysis; e) humidity profile feature sequence: vertical humidity distribution features compressed by wavelet analysis; f) rapid enhancement label: a binary label indicating an intensity increase of ≥7 m / s over the next 12 hours.

[0067] The multimodal spatiotemporal sequence dataset x represents a sequence sample (containing N consecutive time steps), Represents the i-th sample (sequence), which contains the features of n time steps. Its format is: <{remote sensing cloud image 1, ..., remote sensing cloud image n},{Typhoon intensity 1, ...,Typhoon intensity n},{Uprush cloud top brightness temperature 1, ...,Uprush cloud top brightness temperature n}, {temperature profile feature 1,...,temperature profile feature n},{Humidity profile feature 1, ...,Humidity profile feature n}, Quick Enhancement Markup>. represents the i-th typhoon observation sequence sample, which contains a feature sequence of n consecutive time steps (with an interval of 6 hours) 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 a deep feature vector extracted from the remote sensing cloud image through the convolutional neural network (CNN) encoding pre-trained with the basic cloud image dataset. Based on the ResNet-50 pre-training model, its data form is the original satellite image (Himawari-8 / FY4), a 512-dimensional vector; the data form of the typhoon intensity sequence is the central maximum wind speed (m / s), a scalar; the data form of the overburden cloud top brightness temperature sequence is the infrared brightness temperature value (K), which is a scalar obtained by standardization; the temperature profile feature sequence is the vertical temperature distribution curve, the feature vector obtained by compression of the Daubechies wavelet basis function, a 128-dimensional vector; the humidity profile feature sequence is the vertical humidity distribution curve, the feature vector obtained by compression of the Daubechies wavelet basis function, a 128-dimensional vector; the rapid enhancement mark 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] Cubic spline interpolation is used to achieve spatiotemporal alignment of multi-source observation data;

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

[0071] Auxiliary dataset: Typhoon life cycle identifier (1 indicates rising period, 0 indicates falling period) and short-term intensity difference parameter (whether the intensity difference between time t and the previous 3 / 6 / 9 hours meets the standard), which are time-aligned with the multimodal spatiotemporal series data. The format is <life cycle identifier, short-term intensity difference parameter>.

[0072] S120: The unified manifold approximation and projection (UMAP) method is used to preprocess the high-dimensional multimodal spatiotemporal series data to reduce the dimensionality and generate a one-dimensional time series embedding representation of the typhoon observation series;

[0073] In order to fully extract the potential spatiotemporal pattern information from multi-source heterogeneous features such as remote sensing cloud images, typhoon intensity sequences, overshoot cloud top characteristics, and temperature and humidity profiles, step S120 is used to implement multimodal data preprocessing. The unified manifold approximation and projection (UMAP) method is used to reduce the dimensionality of high-dimensional multimodal spatiotemporal series data for subsequent time series modeling and rapid enhanced classification.

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

[0075] First, in the high-dimensional feature space, a weighted adjacency graph that preserves the local structure is constructed in the high-dimensional space. The sample similarity is calculated based on the weighted Euclidean distance and the local neighborhood scale parameter. The local neighborhood scale parameter is negatively correlated with the rate of change of typhoon central pressure or the intensity of wind speed mutation, which is used to enhance the 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 and other information). : The multimodal feature vector at the jth time step, used to compare Compare similarities. : and The weighted Euclidean distance between them is used to measure the differences in the typhoon development status at different time points in the multimodal fusion space. : The local neighborhood scale parameter of the i-th time step, used to construct When the adjacency graph is centered, the scale of its local neighborhood is adjusted. Its value is usually negatively correlated with the central pressure change rate or wind speed mutation intensity of the typhoon at time step i. The smaller the value, the The smaller the local neighborhood around the point, the better The more sensitive the 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 between the two time steps reflects whether the typhoon trends at two time steps are close or continuous. This score is used as the edge weight for constructing the high-dimensional space adjacency graph.

[0078] This similarity function introduces a local Gaussian kernel mechanism, which enables the model to perceive the nonlinear local disturbance structure during the rapid intensification of typhoons and is inspired by the idea of ​​hyperbolic space geometry.

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

[0080] In the low-dimensional embedding space, define the similarity between point pairs , to preserve the structural information of the original space:

[0081]

[0082] in, is the Euclidean distance, which is used to measure the evolution trajectory of the typhoon development state in low-dimensional space. : The low-dimensional embedding representation of the i-th time step, which is a one-dimensional scalar, represents the embedding characteristics of the comprehensive situation of the typhoon at that time point. : The low-dimensional embedding representation of the j-th time step, which participates in similarity modeling as a neighbor of i. : The normalized low-dimensional space similarity score reflects whether the embedded space retains the proximity relationship in 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 has the ability to resist outliers and gradient explosions, and is suitable for non-Euclidean time series embedding modeling. It helps to retain the geometric neighborhood information of the "typhoon intensity turning point" and "rapid intensification trend", and can enhance the model's ability to characterize the rapid intensification stage of extreme typhoons.

[0084] Step S123: similarity is maintained and optimized;

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

[0086] ;

[0087] in, : An overall loss function that measures the degree to which the embedded sequence, after dimensionality reduction, retains the original typhoon multimodal observation structure. The smaller the loss, the more accurately the low-dimensional sequence can represent the similarity between 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), reflecting the 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), which reflects whether the time embedding after dimensionality reduction maintains the original state relationship. : It represents the penalty term of the original similarity information for the deviation of low-dimensional similarity, which is used to strengthen structure preservation. : It is used to suppress the incorrect aggregation of originally "distant" time points by low-dimensional embedding and improve the distinction between non-similar points. : Different observation time points in the time series (such as typhoon center position, intensity level, wind field distribution and other information at different times) meet Used to avoid double counting.

[0088] The optimization process is iterated using the Adam optimizer, and finally the typhoon development time series embedding with low dimension but good structural consistency is obtained, that is, the one-dimensional time series embedding representation :

[0089]

[0090] in That is, it is the embedded representation of the typhoon observation sequence that integrates multimodal features at the tth time step. Through the above-mentioned UMAP dimensionality reduction method, 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 modeling and rapid enhanced prediction.

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

[0092] Among them, the typhoon observation network model is a multimodal spatiotemporal fusion deep learning architecture, whose core is composed of a variational attention recurrent neural network (VA-RNN), and hyperparameter optimization is performed through the improved Harris Hawk Optimization algorithm (DDHHO).

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

[0094] For each time step t, the auxiliary data input vector : [life cycle identifier, short-term intensity difference parameter] = [life cycle identifier_t, Δintensity3h_t, Δintensity6h_t, Δintensity9h_t], where life cycle identifier_t: typhoon life cycle identifier at moment t, 1 indicates rising period, 0 indicates falling period; Δintensity3h_t, Δintensity6h_t, Δintensity9h_t: intensity difference between moment t and the previous 3, 6, and 9 hours (or a binary variable, indicating whether the standard is met). Optional, short-term intensity difference parameters are three binary variables (0 / 1), indicating whether the standard is met.

[0095] The variational attention recurrent neural network is configured with independent input channels: a) the first channel inputs the reduced-dimensional time series embedding sequence; b) the second channel inputs the auxiliary data;

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

[0097] ;

[0098] in, They represent the output of the first type of attention head, the first type of attention head, and the first type of attention head respectively; is the internal weight coefficient (specifically, it can be 0.6); Represents the one-dimensional time series embedded feature sequence (shape is (shape is) in the sliding window of length L (Window_Length) with the time step as the center (or end) at time step t ), is its characteristic dimension; To express Strictly time-aligned, auxiliary data feature window sequence within the same window length L (also in shape ), is its characteristic dimension; is the dynamic gating coefficient, and its value is adaptively adjusted according to the auxiliary data state:

[0099] When the auxiliary data detects an intensity transition When the gating coefficient , strengthen the cross-attention effect;

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

[0101] In an embodiment of the present invention, the auxiliary data includes: a typhoon life cycle identifier, which is encoded as a binary scalar; and a multi-scale intensity difference parameter, which is encoded as a one-hot vector. Specifically:

[0102] Auxiliary data feature vector At time step t, it contains: typhoon life cycle identifier (1-dimensional scalar, 1 = rising period, 0 = falling period), as well as Δ intensity 3h_t, Δ intensity 6h_t, Δ intensity 9h_t. Therefore, is L consecutive A sequence of shapes [L, 4].

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

[0104] The training process of the typhoon observation network model uses a variational attention mechanism recurrent neural network (VA-RNN) to process one-dimensional time series embedding representation 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 a query vector, a key vector, and a value vector through linear transformation to capture the correlation of key areas 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] ; : are the query, key, and value vector matrices, which are linear transformations of features at different times or in different modalities; : The dimension of the key vector, used for scaling and to ensure numerical stability.

[0108] This attention mechanism automatically identifies important spatiotemporal feature patterns that are relevant to future rapid enhancements.

[0109] Step S1322: Using a parallel multi-head structure to fuse multi-scale features: Each attention head captures signals from different physical mechanisms, such as strong convective cloud characteristics or sudden changes in temperature and humidity profiles; the multi-head outputs are spliced ​​and fused across heads using a linear mapping matrix to obtain a fused feature map;

[0110] The multi-head attention mechanism includes three specific types of attention heads. Each type consists of multiple parallel heads. When generating query (Q), key (K), and value (V) vectors, it includes:

[0111] (1) The multi-head attention mechanism includes three types of dedicated attention heads:

[0112] The first type of attention head (time series embedding head): a total of Mz, only receives one-dimensional time series embedding feature window sequence , whose query (Q), key (K), and value (V) vectors are represented by a dedicated weight matrix (Initialization follows 𝒩(0,0.1)) Projection independently The feature vector of each time step is generated to capture the mutation characteristics of the typhoon development process;

[0113] The second type of attention head (auxiliary data head): a total of Ma heads, only receiving auxiliary data feature window sequence , for which query (Q), key (K), value (V) vectors are represented by a dedicated weight matrix (Initialization follows 𝒩(0,0.05)) Projection independently The feature vector of each time step is generated to monitor the intensity transition signal;

[0114] The third type of attention head (cross-modal head): a total of Mc heads, which simultaneously receive a one-dimensional time series embedded feature window sequence and auxiliary data feature window sequence , whose query (Q) vector is represented by the weight matrix Projected separately The feature vector of each time step is generated; its key (K) and value (V) vectors are calculated through the weight matrix Projected separately , the feature vector of each time step is generated to capture the cross-pattern of the two types of features;

[0115] The multi-head mechanism helps comprehensively capture key enhancement signals at different scales and regions. Using three types of attention heads, the model adaptively determines which time steps rely more heavily on time series embedding features (for example, when a typhoon is stable) and which time steps rely more heavily on auxiliary data (for example, when an intensity jump is detected). This approach preserves the multi-head attention mechanism of the original solution while enhancing the model's ability to integrate multi-source data.

[0116] Concatenate the outputs head_m (m=1, ..., M) of all the above M = Mz + Ma + Mc attention heads:

[0117] ;

[0118] in, : the output of the mth attention head, focusing on a specific time window, spatial region or physical variable (such as severe convective cloud area, wet layer 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 splicing operation; Represents the fused feature map output by multiple attention mechanisms.

[0119] Step S1323: Perform temporal window compression aggregation on the fusion feature map, and generate window aggregation features by global average pooling along the time dimension ;

[0120] set up is the multi-head attention output tensor corresponding to time step t, and its shape is Batch is the number of batch samples, L is the input window length (number of time steps), is the feature dimension of the attention output, and the output dimension of each attention head is , then the feature dimension after splicing is ; Perform global average pooling over time on the window sequence of each sample to generate window aggregation features , used for subsequent RNN sequence modeling:

[0121] ;

[0122] in, It represents the aggregate feature of the entire window (including L time steps) ending at time step t, which encodes the comprehensive evolution state of the typhoon in the window [t-L+1, t]; , represents the feature vector of the multi-head attention output tensor at the τth position in the window sequence corresponding to time step t (i.e. For time steps t < L, a zero-padding method is used to construct a complete window.

[0123] Step S1324: Input the window aggregation features into a recurrent neural network to model 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 probability of rapid enhancement in the next 12 hours.

[0124] The window aggregated feature sequence actually contains features from time steps L to T (where T=N, which is the same as the length of the original observation sequence) Input the recurrent neural network (RNN) in chronological order to model the temporal evolution and capture the dynamic change law of time:

[0125] ;

[0126] : Multimodal embedded feature representation of the typhoon at time t (such as fused wind speed field, sea temperature, pressure structure and other information). : The hidden state vector at time t, encoding 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 unit, used to model the temporal dependence characteristics of typhoon development. During training, it is necessary to save the hidden state at each moment .

[0127] Finally, the last moment will be hidden Input to a classifier (fully connected layer + Sigmoid) output fast enhancement probability :

[0128] ;

[0129] in, : The global time series representation of the typhoon at the last moment (i.e., the moment before the forecast target), integrating all previous observation states. : The weight matrix of the classifier, used to linearly map the hidden state to the prediction space. : Bias term of the classifier. : Sigmoid function, used to compress the linear output to the [0,1] interval, indicating the predicted probability of rapid intensification of the typhoon sample. : The confidence level of the model’s prediction of rapid enhancement. The closer the value is to 1, the higher the possibility of rapid enhancement.

[0130] The classifier is composed of a fully connected layer and a Sigmoid activation function, and outputs a probability value; and the probability judgment threshold is set to 0.55. , then the typhoon sample is judged to be "rapidly intensifying".

[0131] The total loss function of model training consists of a binary cross entropy loss function and a loss function containing feature reconstruction constraints. The total loss function is:

[0132] ;

[0133] in: To reconstruct the loss weight coefficient, it can be dynamically adjusted according to the typhoon type, with a value range of 0.05 to 0.2. When a typhoon eyewall reorganization event is detected, It can be temporarily increased to 0.25 to strengthen the feature preservation constraint. Reconstruction loss function :

[0134] ;

[0135] is a reconstructed mapping function that satisfies: , The dimension is , is the hidden state dimension, Embed dimension for one-dimensional time series; is the mapping function of the fully connected layer, which transforms the hidden state Mapped to a one-dimensional time series embedding representation of the t-th time step Same dimensions; is the layer normalization function; the reconstruction target is the embedding of the window end time t .

[0136] Binary cross entropy loss for:

[0137] ;

[0138] in, : Real fast enhanced labeling; : Model prediction confidence.

[0139] Step S133: Optimizing the typhoon observation network model;

[0140] In order to improve the generalization performance and accuracy of the network based on spatiotemporal sequence and multimodal feature fusion in the typhoon rapid intensification prediction task, 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 common problems of slow convergence and easy falling into local optimality in traditional methods.

[0141] Specifically, the typhoon observation network model optimization process uses the improved Harris Hawks optimization algorithm (DDHHO) to optimize hyperparameters, including:

[0142] Step S1331: Initialize the population using Tent chaotic mapping combined with random perturbations;

[0143] In the traditional HHO algorithm, the optimized population initialization uses a random uniform distribution, which easily causes individuals to cluster in local areas of the typhoon sample feature space, affecting the global modeling of the coupled relationship between multimodal complex features (such as cloud texture, sea temperature distribution, and air pressure changes). DDHHO uses the Tent chaos map combined with random factors to generate the initial population, enhancing distribution diversity:

[0144] ; : The current position of the nth individual in the Tent map, indicating 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 ergodicity; :exist The random perturbation top sampled within the interval increases the uncertainty and diversity of the initial population; N: The population size in the optimization algorithm, which indicates the number of candidate individuals participating in the hyperparameter search of the typhoon prediction model.

[0145] This formula is used to generate the chaotic sequence of initial optimized individuals , to improve the distribution diversity of the population in the complex multimodal feature space. By introducing tent mapping and random perturbations, the optimized population can more fully cover the complex combination space of multimodal features such as cloud texture, sea temperature, and air pressure, avoiding falling into local optimality.

[0146] After generating chaotic individuals, the specific hyperparameter values ​​are obtained through inverse mapping:

[0147] ;

[0148] : The model hyperparameter values ​​corresponding to the nth individual, such as the dimension of the loop layer, the number of attention heads, the size of the convolution kernel, etc. : The minimum value range of this hyperparameter, indicating the starting point of the model's minimalist structure; : The maximum value range of this hyperparameter represents the upper bound of the model's complexity; : Individual position values ​​processed by tent mapping after normalization, controlling the mapping weight.

[0149] This formula converts the chaotic mapping result Mapped to specific model hyperparameter values , used to control the network structure and training configuration. Mapping can generate a variety of network structure configurations for training the "Typhoon Rapid Intensification Prediction Model" and iteratively search for the optimal structure in the DDHHO framework to improve the model's perception and generalization capabilities of modal coupling changes during typhoon evolution.

[0150] Step S1332: Design a nonlinear escape energy function to dynamically control the ratio of global exploration to local development;

[0151] Figure 3 Schematic diagram of the prediction model optimization process of the embodiment of the present invention. Figure 3 As shown in Figure 1, DDHHO performs a nonlinear reconstruction of the escape energy E to simulate the characteristic that uncertainty in typhoon development gradually decreases over time. The escape energy E controls the ratio of global and local searches, and its update formula is as follows:

[0152] ; r: the current iteration number, indicating the time process of the algorithm evolution; : The maximum number of iterations, indicating the end time of the optimization process; : Nonlinear mapping adjustment factor, used to control the slope and directionality of the escape energy curve, reflecting the stage-by-stage 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, and is used to simulate nonlinear uncertainties in the typhoon intensification process (such as development mutations or energy release hysteresis).

[0153] This nonlinear mapping aims to achieve a smooth transition from early high exploratory nature (E≈1, emphasizing global modeling of large-scale relationships between multimodal information) to later high exploitative nature (E≈0, improving local convergence capabilities and integrating high-confidence feature representations), effectively adapting to the shift in decision-making focus from breadth search to depth mining during the rapid intensification and evolution of typhoons.

[0154] Step S1333: Search strategy optimization;

[0155] (1) Optimization of the global exploration phase

[0156] When the escape energy E≥1, the Cauchy mutation mechanism is used to update the parameters for global exploration optimization to enhance the ability to escape from the local optimum;

[0157] In response to the large amount of noise and inconsistency in the multimodal feature space (such as cloud cover interference in remote sensing images), DDHHO introduces a Cauchy mutation mechanism to perturb the eagles individually, increasing the probability of escaping the local optimum:

[0158] ;

[0159] : The parameter vector represented by the current individual, indicating the parameter position of a candidate optimization 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 (for example, areas with more drastic changes in image texture require greater perturbation), ,wherein, the cloud interference index is calculated based on the proportion of cloud cover area in the satellite image; :The standard Cauchy distribution samples have the characteristics of thick tails, allowing larger step-size jump disturbances, broadening the search radius, and helping to break through the feature traps caused by cloud occlusion, local anomalies, etc. in remote sensing images.

[0160] This strategy enhances global exploration and jumping capabilities, avoiding being trapped in local optima in complex multimodal feature spaces. It is particularly suitable for modeling typhoons with sudden structural changes or extreme intensification. Combining chaotic initialization with nonlinear escape energy design, this mechanism strengthens the model's ability to identify sudden changes in typhoon characteristics during sudden intensification, improves its robustness to scenarios such as remote sensing interference and high-latitude disturbances, and enhances the fitting accuracy of feature fusion representations.

[0161] (2) Optimization of local development stage

[0162] When the escape energy E<1, a combined mechanism of adaptive inertia weight and random perturbation is used to perform local development optimization and enhance fine-grained feature response.

[0163] In order to enhance the modeling of local details (such as the response of pressure gradient and wind speed difference at small scale), a combined mechanism of adaptive inertia weight and random perturbation is introduced. The inertia weight adjustment formula is:

[0164] ;

[0165] : Inertia weight, which is used to control the update amplitude and inertia trend of the current individual parameters, and affects the fitting accuracy of small-scale meteorological characteristics (such as local pressure changes and wind speed differences) during the optimization process; : The upper limit of the inertia weight is usually set to a larger value to quickly converge to the potential area in the initial optimization process and ensure that the model can cover the global characteristic distribution in the typhoon multimodal data; : The lower limit of the inertia weight, set to a smaller value, is used to optimize the local sensitivity to key enhancement signals in the later stage and refine microstructures such as "typhoon eyewall reorganization" or "deep convective outbreak"; r: The current iteration number, representing the time node in the model learning process; : The maximum number of iterations, which determines the total optimization step length and inertia decay rate.

[0166] Among them, the upper and lower limits of the adaptive inertia weight are associated with the typhoon type: Oceanic typhoon: ; Landfall typhoon: , used to adapt to the sensitivity of microscopic characteristics of different typhoon intensification processes.

[0167] This formula dynamically adjusts the inertia strength of search individuals during the update process, enabling a smooth transition from global exploration to local exploitation. It is particularly well-suited for capturing microscopic disturbance signals during typhoon intensification. A large inertia weight ensures a more exploratory model in the initial stages, accelerating convergence; a small inertia weight allows for later focus on detailed modeling and identifying boundary samples, such as short-term wind pressure anomalies before a typhoon suddenly intensifies.

[0168] The random disturbance in the local development stage adopts Gaussian distribution:

[0169] ;

[0170] in, is the global optimal solution found by the rth iteration; is the solution vector of the rth iteration. is a random noise vector, , multivariate normal distribution, mean is 0, covariance matrix is ; is a unit matrix to ensure that each dimension of the noise is independent; the standard deviation σ is negatively correlated with the strength of the weak signal that precedes the typhoon strengthening, and the specific value is σ=0.1×(1-the recall rate of the current optimal model on the validation set).

[0171] During the local search phase, a random perturbation joint mechanism is introduced to slightly perturb the current individual, aiming to escape the local minimum trap and enhance the model's responsiveness to true rapid intensification signals. This mechanism controls whether to introduce perturbation jumps when the individual converges to a local optimum, reducing the risk of "false intensification" or "weak signal misinterpretation" and improving the model's robustness to true rapid intensification. The local perturbation mechanism is particularly useful in the fuzzy pre-intensification phase, where probabilistic perturbations stimulate the model's search activity in boundary regions, thereby identifying potential signs of explosive intensification (such as unusual cloud accumulation near high sea temperature areas).

[0172] The DDHHO optimization strategy comprehensively considers the nonlinear, multi-scale, and multi-modal evolution characteristics of the rapid intensification of typhoons, effectively improving the quality of the model's hyperparameter configuration, thereby significantly enhancing the prediction model's fitting ability and discrimination accuracy for complex meteorological processes.

[0173] 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. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the methods.

[0174] It should also be noted that, in the embodiments of the present application, relational terms such as first and second, etc. are merely used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device comprising the elements.

[0175] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined in the embodiments of the present application may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown in the embodiments of the present application, but rather will conform to the widest scope consistent with the principles and novel features disclosed in the embodiments of the present application.

Claims

1. A typhoon rapid intensification prediction method based on the fusion of spatiotemporal series and multimodal features, characterized by: The following steps are involved: Based on typhoon optimal path data and multi-source satellite observation data, a multimodal spatiotemporal series dataset and auxiliary dataset are constructed; The unified manifold approximation and projection method is used to preprocess the high-dimensional multimodal spatiotemporal series data to reduce the dimensionality and generate a one-dimensional time series embedding representation of the typhoon observation series. The one-dimensional time series embedding representation and auxiliary data are used as independent input channels and input into a trained typhoon observation network model to predict the probability of rapid intensification of the typhoon; Among them, the typhoon observation network model is a multimodal spatiotemporal fusion deep learning architecture, whose core is composed of a variational attention recurrent neural network, and hyperparameter optimization is performed through an improved Harris Hawks optimization algorithm.

2. The typhoon rapid intensification prediction method based on the fusion of spatiotemporal series and multimodal features according to claim 1 is characterized in that: The steps for constructing the multimodal spatiotemporal sequence dataset are as follows: N consecutive typhoon observation sequences are collected at 6-hour intervals. Each sequence contains: a) remote sensing cloud image sequence: deep feature vectors extracted from remote sensing cloud images using a convolutional neural network pre-trained with a basic cloud image dataset; b) typhoon intensity sequence: maximum central wind speed values; c) upthrust cloud top brightness temperature sequence: infrared brightness temperature measurements; d) temperature profile feature sequence: vertical temperature distribution features compressed by wavelet analysis; e) humidity profile feature sequence: vertical humidity distribution features compressed by wavelet analysis; f) rapid enhancement flag: a binary label indicating an intensity increase of ≥7 m / s over the next 12 hours. Cubic spline interpolation is used to achieve spatiotemporal alignment of multi-source observation data; Data augmentation via path simulation and feature perturbation.

3. The typhoon rapid intensification prediction method based on the fusion of spatiotemporal series and multimodal features according to claim 2 is characterized in that: The specific process of the dimensionality reduction preprocessing includes: A weighted adjacency graph that preserves local structure is constructed in high-dimensional space, and sample similarity is calculated based on weighted Euclidean distance and a local neighborhood scale parameter; wherein the local neighborhood scale parameter is negatively correlated with the rate of change of typhoon central pressure or the intensity of wind speed mutation, and is used to enhance sensitivity to precursors of rapid intensification; Define a similarity function against outliers in 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 the one-dimensional time series embedding representation is generated by iterative optimization.

4. The typhoon rapid intensification prediction method based on the fusion of spatiotemporal series and multimodal features according to claim 3 is characterized in that: in, The training process of the typhoon observation network model uses a recurrent neural network with a variational attention mechanism to process the one-dimensional time series embedding representation and the auxiliary data, specifically including: Inputting the one-dimensional time series embedding representation and the auxiliary data into a multi-head attention mechanism, generating a query vector, a key vector, and a value vector through linear transformation to capture the correlation of key regions in different time steps; A parallel multi-head architecture is used to fuse multi-scale features: each attention head captures signals from different physical mechanisms, such as strong convective cloud characteristics or sudden changes in temperature and humidity profiles; multiple head outputs are spliced ​​and fused across heads using a linear mapping matrix to produce a fused feature map. Perform temporal window compression aggregation on the fused feature map, and generate window aggregation features by global average pooling along the time dimension; The window aggregation features are input into a recurrent neural network to model 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 probability of rapid enhancement in the next 12 hours.

5. The typhoon rapid intensification prediction method based on the fusion of spatiotemporal series and multimodal features according to claim 4 is characterized in that: The classifier is composed of a fully connected layer and a Sigmoid activation function, and outputs a probability value; and the probability judgment threshold is set to 0.

55. When the probability value exceeds 0.55, it is judged as rapid enhancement; The total loss function of model training consists of the binary cross entropy loss function and the time series embedding reconstruction loss function.

6. The typhoon rapid intensification prediction method based on the fusion of spatiotemporal series and multimodal features according to claim 5 is characterized in that: The multi-head attention mechanism of the variational attention recurrent neural network includes three types of specialized attention heads: The first type of attention head: a total of Mz, only receives one-dimensional time series embedding feature window sequence , used to capture the sudden change characteristics during the development of typhoons; The second type of attention head: a total of Ma heads, only receiving auxiliary data feature window sequence , used to monitor the intensity transition signal; The third type of attention head: a total of Mc heads, which simultaneously receive a one-dimensional time series embedded feature window sequence and auxiliary data feature window sequence , used to capture the cross-patterns of the two types of features.

7. The typhoon rapid intensification prediction method based on the fusion of spatiotemporal series and multimodal features according to claim 6 is characterized in that: The total loss function of the model training is: in: is the time series embedding reconstruction loss weight coefficient, ranging from 0.05 to 0.2; the time series embedding reconstruction loss function : is a reconstructed mapping function that satisfies: , The dimension is , is the hidden state dimension, Embed dimension for one-dimensional time series; is the mapping function of the fully connected layer, which transforms the hidden state Mapped to a one-dimensional time series embedding representation of the t-th time step Same dimensions; is the layer normalization function; Binary cross entropy loss for: in, : Real fast enhanced labeling; : Model prediction confidence.

8. The typhoon rapid intensification prediction method based on the fusion of spatiotemporal series and multimodal features according to claim 7 is characterized in that: in, The auxiliary data includes: a typhoon life cycle identifier, encoded as a binary scalar; and a multi-scale intensity difference parameter, encoded as a one-hot vector.

9. The typhoon rapid intensification prediction method based on the fusion of spatiotemporal series and multimodal features according to claim 8 is characterized in that: in, The variational attention recurrent neural network dynamically integrates dual-channel outputs through a gated fusion mechanism: in, They represent the output of the first type of attention head, the first type of attention head, and the first type of attention head respectively; is the internal weight coefficient; is a dynamic gating coefficient whose value is adaptively adjusted according to the auxiliary data state: when the auxiliary data detects an intensity transition When the gating coefficient ; When the auxiliary data shows a stable state, .

10. The typhoon rapid intensification prediction method based on the fusion of spatiotemporal series and multimodal features according to claim 1, characterized in that: Hyperparameter optimization is performed using the improved Harris Hawks optimization algorithm, including: Use Tent chaotic map combined with random perturbations to initialize the population; Design a nonlinear escape energy function to dynamically control 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 optimization to enhance the ability to escape from the local optimum; When the escape energy E<1, a combined mechanism of adaptive inertia weight and random perturbation is used to perform local development optimization and enhance fine-grained feature response.

Images