A typhoon wind speed prediction method based on Mamba-ASPP and R-TCN

Abstract

The present invention discloses a typhoon wind speed prediction method based on Mamba-ASPP and R-TCN, and the steps are as follows: obtaining a satellite cloud image data set and performing preprocessing; constructing a Mamba-ASPP module for spatial feature extraction, constructing an R-TCN module for temporal feature extraction, obtaining a global spatial feature map of an input image and a temporal feature map splicing, placing an ITSAM module for comprehensive feature processing, and obtaining a predicted value of the typhoon wind speed; integrating the three modules into an initial wind speed prediction model, optimizing and training the initial wind speed prediction model hyperparameters with a seagull algorithm optimized by an adaptive genetic algorithm, and obtaining a target wind speed prediction model; inputting the image data of the previous hour into the target wind speed prediction model, and predicting the typhoon wind speed of the next hour. The method of the present invention effectively improves the efficiency and accuracy of typhoon prediction by integrating feature information of different time and space scales to cope with the complexity of typhoon wind speed changes.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological forecasting, and in particular to a typhoon wind speed forecasting method based on Mamba-ASPP and R-TCN. Background Art

[0002] Accurate prediction of typhoon wind speed is crucial in meteorology. As a strong tropical cyclone, typhoon has a complex spatial structure and drastic temporal variation characteristics, making accurate prediction of its wind speed and intensity a major difficulty in scientific research.

[0003] Traditional typhoon wind speed prediction methods mainly rely on numerical weather forecast models, which are based on mathematical simulations of physical processes and can better reflect the dynamic and thermodynamic processes of typhoons. By improving the physical parameterization scheme of numerical models, increasing grid resolution and improving initialization techniques, the prediction ability of numerical weather forecast models has gradually increased. However, due to the complexity and variability of typhoon structures, traditional numerical models still face problems such as large computational complexity, high time overhead and large uncertainty when making high-precision wind speed predictions, especially when facing rapidly changing typhoon processes, the accuracy and timeliness of predictions are still insufficient. With the advancement of science and technology, meteorological observation methods are constantly enriched, such as satellite remote sensing, ocean buoys, radar observations, etc., which can provide rich typhoon observation data, which provides more detailed data information for typhoon wind speed prediction. The fusion technology of multi-source data has improved the timeliness and accuracy of typhoon intensity monitoring, but how to mine the features that have a significant impact on typhoon wind speed changes from these massive data is still a problem to be solved.

[0004] In recent years, the development of artificial intelligence and deep learning technology has brought new opportunities for typhoon wind speed prediction. Deep learning methods can automatically extract spatial and temporal features from a large amount of observation data and establish nonlinear relationship models, thereby improving the accuracy and efficiency of predictions. However, when predicting typhoon wind speed, existing deep learning methods often fail to fully combine the global structural information and local detail information of the typhoon, and it is difficult to cope with the complexity of typhoon wind speed changes. In addition, how to effectively integrate feature information at different time and spatial scales is also one of the challenges faced by deep learning models in typhoon wind speed prediction. Summary of the invention

[0005] The problem to be solved by the present invention is to provide a typhoon wind speed prediction method based on Mamba-ASPP and R-TCN, which can cope with the complexity of typhoon wind speed changes by integrating characteristic information of different time and space scales, and improve the efficiency and accuracy of typhoon prediction.

[0006] The present invention adopts the following technical solution: a typhoon wind speed prediction method based on Mamba-ASPP and R-TCN, comprising the following steps:

[0007] Step 1: Obtain a satellite cloud image dataset and preprocess the image to obtain a preprocessed image dataset;

[0008] Step 2: construct a Mamba-ASPP module, input the image data of adjacent frames in the preprocessed image data set into the Mamba-ASPP module, perform spatial feature extraction, and obtain a global spatial feature map of the input image;

[0009] Step 3, construct an R-TCN module, input the image data of adjacent frames in the preprocessed image data set into the R-TCN module, perform temporal feature extraction, and obtain a global temporal feature map of the input image;

[0010] Step 4: construct an ITSAM module, splice the global spatial feature map and the global temporal feature map of the input image, put them into the ITSAM module for comprehensive feature processing, and obtain the predicted value of the typhoon wind speed;

[0011] Step 5: Integrate the constructed Mamba-ASPP module, R-TCN module, and ITSAM module to form an initial wind speed prediction model, optimize and train the hyperparameters of the initial wind speed prediction model using the Seagull algorithm optimized by the adaptive genetic algorithm, and use the trained model as the target wind speed prediction model;

[0012] Step 6: Input the image data of the previous hour into the target wind speed prediction model to predict the typhoon wind speed of the next hour.

[0013] Preferably, the image is preprocessed in step 1, including the following sub-steps:

[0014] Step 101: Use Gaussian filtering to remove image noise from the acquired satellite cloud image data set;

[0015] Step 102, performing threshold segmentation on the image using an adaptive threshold segmentation method;

[0016] Step 103, use 3 The principle is to identify outliers in the image and replace them with the median of adjacent pixels, and use bilinear interpolation to handle missing values;

[0017] Step 104: Use the Min-Max normalization method to normalize the image after bilinear interpolation, and map the pixel values ​​to the range of [0, 1].

[0018] Preferably, in step 2, constructing a Mamba-ASPP module and performing spatial feature extraction includes the following sub-steps:

[0019] Step 201, constructing a Mamba-ASPP module, combining a Mamba network module and an ASPP network module, in the Mamba network module, extracting spatial features of the image through a state space module and a depthwise separable convolution, and in the ASPP network module, extracting global spatial features of the image through multi-scale dilated convolution;

[0020] Step 202: The pre-processed image is input into the Mamba-ASPP module. The image is sequentially processed Depthwise separable convolutional layers, Depthwise separable convolutional layers, Depthwise separable convolutional layers and Depth-wise separable convolutional layer, resulting in four feature maps , , , ;

[0021] Step 203: The feature map With feature map Perform channel splicing and pass through The convolution layer linearly combines the features of each channel, and the linear layer and Relu activation function layer enhance the nonlinear representation ability of the feature map, and then enter the SSM-ASPP module to extract multi-scale spatial features to obtain the primary feature map. ;

[0022] Step 204: feature map The feature map obtained after the linear layer and the Relu activation function layer is the same as the primary feature map Perform channel splicing and enter the SSM-ASPP module again to extract multi-scale spatial features to obtain the intermediate feature map ;

[0023] Step 205: The feature map With feature map To splice channels, The convolution layer linearly combines the features of each channel, passes through the linear layer and the Relu activation function layer, and is combined with the intermediate feature map Perform channel splicing and enter the SSM-ASPP module to extract multi-scale spatial features to obtain high-level feature maps ;

[0024] Step 206: Get the primary feature map , Intermediate feature map , high-level feature maps Perform channel splicing and pass through Convolutional layer, linear layer, and Relu activation function layer are used to obtain the global spatial feature map of the input image.

[0025] Preferably, in step 3, constructing an R-TCN module and performing temporal feature extraction includes the following sub-steps:

[0026] Step 301: construct an N-layer R-TCN network module. Each layer of the R-TCN network module uses causal dilation convolution to model the time series, and fuses the dilated pyramid convolution to obtain the causal dilated pyramid convolution. For the input feature map , the temporal features are extracted through the causal dilated pyramid convolution, and the temporal features calculated by each dilated convolution kernel are spliced ​​to obtain the temporal feature map ;

[0027] Step 302: Time characteristic graph Put it into the first layer of TCN network module, and after the batch normalization layer, Swish activation function layer, and Stochastic Depth regularization mechanism layer, the nonlinear transformation ability and stability of the features are improved to obtain the feature map , the input feature map Through residual connection and feature map Add to get the feature map ;

[0028] Step 303: The feature map As input, it is put into the second layer TCN network module to obtain the feature map , the feature map Through residual connection and feature map After addition, it is put into the third-layer TCN network module as input. After repeating N-1 times, the output of the N-1th layer TCN network module is added to the output of the Nth layer TCN network module through a residual connection to obtain the global temporal feature map of the input image.

[0029] Preferably, in step 4, constructing an ITSAM module and performing comprehensive feature processing includes the following sub-steps:

[0030] Step 401, construct an ITSAM module, including a multi-head attention mechanism, an embedding submodule, a spatial attention submodule, and a temporal attention submodule; the spatial attention submodule is used to weight and select spatial features, and the temporal attention submodule is used to strengthen and select key information in the temporal dimension;

[0031] Step 402: The global spatial feature map and the global temporal feature map of the input image are spliced ​​and input into the embedding submodule for feature fusion, and then input into the spatial attention submodule and the temporal attention submodule for processing respectively. The obtained feature maps are multiplied and input into the multi-head attention layer to fuse feature information from different dimensions.

[0032] Step 403: Add the multiplied feature map to the feature map output by the multi-head attention layer through a residual connection to obtain an output feature map;

[0033] Step 404: The output feature map is subjected to global pooling and a fully connected layer. The global pooling performs dimensionality reduction processing on the output feature map and compresses the feature information into a feature vector of a fixed length. The fully connected layer performs linear mapping on the feature vector and converts it into a specific predicted value, which is the predicted value of the typhoon wind speed.

[0034] Preferably, in step 5, the hyperparameters of the initial wind speed prediction model are optimized and trained using the Seagull algorithm optimized by the adaptive genetic algorithm, including the following sub-steps:

[0035] Step 501: Construct a population consisting of different hyperparameter combinations, the population includes m hyperparameter combinations, each of which is represented by a vector , where each element Represents the value of a hyperparameter. A hyperparameter combination has n hyperparameters, each hyperparameter combination is a seagull individual;

[0036] Step 502: select the inverse of the mean square error as the adaptability function, set the maximum number of iterations and the adaptability threshold, put all seagull individuals in the population into the initial wind speed prediction model for training, and calculate the adaptability of each seagull individual through the adaptability function;

[0037] Step 503: Select the top population with the best adaptability value. Seagull individuals, in Randomly select two individuals from and As a parent, use adaptive scale Perform linear crossover to generate two offspring seagull individuals and ;

[0038] Step 504, enter the mutation phase, the mutation operation introduces appropriate randomness and fine-tuning, so that the generated offspring seagull individuals explore new solution space and optimization probability based on the parent generation; for linear crossover to generate two offspring seagull individuals and , each hyperparameter A random number will be generated randomly ,when When , adaptive mutation is performed, otherwise the original value is maintained;

[0039] Two parent seagull individuals and After two offspring individuals are generated through adaptive crossover and mutation, Randomly select two individuals from the seagull individuals and As the parent, two offspring individuals are generated through adaptive crossover and mutation. and ,cycle After that, all the seagull offspring individuals obtained through adaptive crossover and mutation are regarded as the new generation population;

[0040] Step 505: Each seagull individual in the new generation population completes the migration process and the attack process to the target. The migration process simulates the behavior of the seagull approaching the target position, so that the individuals in the population gradually approach potential excellent individuals; the attack process simulates the diving behavior of the seagull when hunting, adjusts the individual position, and makes the individual conduct local search while approaching the target position, thereby enhancing the approach to the optimal individual.

[0041] Step 506: After each seagull individual has gone through the migration process and the attack process, a better new population is generated. After all individuals in the new population are put into the wind speed prediction model for training, the adaptive function is used to calculate the adaptive value of each individual in the new population, and it is determined whether the adaptive value of the best individual in the new population has reached the maximum threshold. If not, the number of iterations is increased by 1, and steps 503 to 506 are repeated to perform adaptive crossover and mutation on the new population. If the adaptive value of the best individual in the population has reached the maximum threshold, Or when the maximum number of iterations is reached, the optimal individual is output and the hyperparameter combination of the optimal individual is obtained.

[0042] Preferably, in step 6, the image data of the previous hour is input into the target wind speed prediction model to predict the typhoon wind speed of the next hour, including the following sub-steps:

[0043] Step 601: Put the obtained optimal individual hyperparameter combination into the initial wind speed prediction model for training. After the training is completed, the optimal weight of the model is obtained, and the target wind speed prediction model is constructed using the optimal weight to perform typhoon wind speed prediction;

[0044] Step 602: Put the satellite cloud image data of the previous hour into the target wind speed prediction model to obtain the cloud image feature map of the next hour, and pass the cloud image feature map through a fully connected layer to obtain the typhoon wind speed prediction value of the next hour. .

[0045] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0046] 1. The typhoon wind speed prediction method of the present invention decomposes the typhoon wind speed prediction into two subtasks: time feature extraction and spatial feature extraction. The R-TCN module is used to perform multi-scale recursive modeling on the time series to capture the dynamic change trend of the typhoon wind speed. At the same time, the SSM-ASPP structure in the Mamba-ASPP module is used to realize dynamic multi-scale convolution to extract complex spatial features, realize adaptive modeling and efficient fusion of time and space features, and effectively solve the problem of uneven performance in the time and space dimensions.

[0047] 2. The typhoon wind speed prediction method of the present invention innovatively integrates the spatiotemporal layered attention mechanism and the multi-head attention mechanism in the ITSAM module, processes the temporal and spatial features through hierarchical separation, and allocates adaptive attention weights to each feature layer. The spatiotemporal layered attention mechanism enables the model to focus on the temporal dynamics and spatial structure of the typhoon wind speed respectively, while the multi-head attention mechanism gives each spatiotemporal feature layer a more detailed weight adjustment, thereby ensuring a more balanced weight distribution during information fusion, and being able to more accurately extract and integrate the key spatiotemporal variation characteristics of the typhoon wind speed, significantly improving the accuracy and robustness of the prediction.

[0048] 3. The typhoon wind speed prediction method of the present invention uses the Seagull algorithm optimized by the adaptive genetic algorithm to tune the key hyperparameters of the model, including the number of training rounds, convolution kernels and learning rate. The adaptive genetic algorithm provides global search capabilities and adaptive adjustment mechanisms, and the Seagull algorithm further performs local fine search to achieve evolutionary optimization of hyperparameters. During the model training process, the hyperparameters are dynamically adjusted through evolutionary optimization to ensure that the model achieves the best performance in the data set and environment, significantly improving the robustness and generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flow chart of the typhoon wind speed prediction method based on Mamba-ASPP and R-TCN of the present invention;

[0050] Figure 2 It is a schematic diagram of the structure of the Mamba-ASPP module of the present invention;

[0051] Figure 3 It is a schematic diagram of the structure of the R-TCN module of the present invention;

[0052] Figure 4 It is a schematic diagram of the ITSAM module structure of the present invention;

[0053] Figure 5 It is the flow chart of AGA-SOA algorithm of the present invention;

[0054] Figure 6 It is a schematic diagram of the SAM module inside the ITSAM module of the present invention;

[0055] Figure 7 It is a schematic diagram of the TAM module inside the ITSAM module of the present invention;

[0056] Figure 8 It is a schematic diagram of the INCM module inside the ITSAM module of the present invention. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the application is further elaborated in detail below in conjunction with the accompanying drawings. The described embodiments are only a part of the embodiments involved in the present invention. All non-innovative embodiments of other researchers in the field on this embodiment belong to the protection scope of the present invention. At the same time, for the step numbering in the embodiment of the present invention, it is only set for the convenience of explanation, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0058] In one embodiment of the present invention, a typhoon wind speed prediction method based on Mamba-ASPP and R-TCN is provided. Figure 1 As shown, the following steps are included:

[0059] Step 1: Obtain a satellite cloud image dataset and preprocess the image to obtain a preprocessed image dataset;

[0060] Step 2: construct the Mamba-ASPP module and put the image data of adjacent frames after preprocessing into the module to extract spatial features;

[0061] Step 3: construct an R-TCN module and input the same pre-processed adjacent frame image data into the module to extract temporal features;

[0062] Step 4: Build an ITSAM module, and combine the feature image results of step 2 and step 3 into the ITSAM module for comprehensive feature processing;

[0063] Step 5: Integrate the three constructed modules to form an initial wind speed prediction model, use the Seagull algorithm optimized by the adaptive genetic algorithm to optimize and train the hyperparameters of the initial wind speed prediction model, and use the trained model as the target wind speed prediction model;

[0064] Step six, input the image data of the previous hour into the target wind speed prediction model to predict the typhoon wind speed of the next hour.

[0065] Specifically, in this embodiment, the satellite cloud image data set obtained in step 1 has 2000 sequences, one sequence has 12 frames of images, one frame every 5 minutes. When inputting the model, these 12 frames of images are used as input to predict the next 12 frames of images. One sequence represents one hour of data. After preprocessing the satellite cloud image data set, the satellite cloud image sequence of the previous hour is input to predict the typhoon wind speed of the next hour.

[0066] Step 1: Preprocessing the image includes:

[0067] Step 101, use Gaussian filtering to remove image noise from the acquired data set, the formula is as follows:

[0068] ;

[0069] in, is the standard deviation of the Gaussian kernel, is the value of the Gaussian kernel, x and y represent the two-dimensional plane coordinates of the Gaussian kernel, The function is a natural exponential function of the form .

[0070] Step 102, using the adaptive threshold segmentation method to perform threshold segmentation on the image, the formula is as follows:

[0071] ;

[0072] in, is the window, N is the number of pixels in the window, C is an adjustment constant, is the coordinate in the image The gray value of the pixel.

[0073] Step 103, use 3 The principle is to identify outliers in the image and replace them with the median of adjacent pixels, and then use bilinear interpolation to handle missing values. The principles and bilinear interpolation formulas are as follows:

[0074] ;

[0075] in, is the standard deviation of the image pixel values, is the image mean, Represents the value of the i-th pixel in the image. If the pixel value in the image satisfies the above formula, it is defined as an outlier.

[0076] ;

[0077] in, Represents the coordinates of the interpolation target point, ,express exist and The relative position between express exist and The relative position between Represents the values ​​of four known neighboring pixels.

[0078] Step 104, use the Min-Max normalization method to normalize the image after bilinear interpolation, and map the pixel values ​​to the range of [0, 1]. The formula is as follows:

[0079] ;

[0080] in, and are the minimum and maximum values ​​of the image pixels, respectively.

[0081] Furthermore, after the image preprocessing is completed, the image data set is divided into a training set and a test set to facilitate the subsequent training in the model, with 80% used as the training set and 20% as the test set.

[0082] Specifically, in this embodiment, step 2 constructs the Mamba-ASPP module and performs spatial feature extraction, such as Figure 2 As shown, including:

[0083] The preprocessed image will first go through The depth-separable convolutional layer extracts the original features of the image, and then passes through Depthwise separable convolutional layers and The depthwise separable convolutional layer expands the receptive field to extract refined features, and finally passes The depthwise separable convolutional layer integrates the features of each channel.

[0084] Preferably, the depthwise separable convolution includes two parts: depthwise convolution and pointwise convolution. Assume that the shape of the input image in this embodiment is ,in are the image length and width, is the number of channels of the input image. The calculation formulas after depth convolution and point convolution are as follows:

[0085] ;

[0086] ;

[0087] in, Represents image coordinates With The eigenvalues ​​at the channels, is the eigenvalue of the input image, Representative in the Channels The weight of , Represents the convolution kernel size. If Depthwise separable convolution, then , All are 3, Point convolution The weight of the convolution kernel, is the number of channels after depthwise convolution, is the index of the number of channels after point convolution. If the number of channels after point convolution is ,but The value ranges from 0 to -1, The output image is at coordinates With The spatial eigenvalues ​​at the channels.

[0088] The input image is sequentially processed through four layers of depth-separable convolution to obtain four feature maps. , , , . The feature map With feature map Perform channel splicing and pass The convolution layer linearly combines the features of each channel, and then goes through the linear layer and the Relu activation function layer to enhance the nonlinear representation ability of the feature map. It then enters the SSM-ASPP module to extract multi-scale spatial features to obtain the primary feature map. .

[0089] Assume that the input feature map entering the SSM-ASPP module is X. First, the feature map X will be processed by the state update formula of SSM.

[0090] SSM combines the current input feature map X and the previous input spatial state feature Combined, the state transfer matrix A and the input mapping matrix B are used to calculate the current spatial state characteristics , and then through the selective gating matrix Control the relative influence of the input and the previous state, so that the state features can adaptively focus on important spatial information and obtain the final state features of SSM , and then pass it into ASPP to extract multi-scale feature maps through convolution operations with different void rates, so as to obtain feature representations at multiple different scales.

[0091] Then, these multi-scale feature maps are combined with the final state features output by SSM spliced ​​together to form a feature map containing multi-scale spatial information. Finally, the spliced ​​feature map is Convolution linear transformation generates the final spatial feature Y. The process formula is as follows:

[0092] ;

[0093] ;

[0094] ;

[0095] ;

[0096] ;

[0097] ;

[0098] in, and is the weight matrix acting on the current input image and the previous input spatial state, is the bias term, is the Sigmoid activation function, represents the element-by-element multiplication of the feature map, is the output eigenvalue under different void ratio r.

[0099] In this embodiment, the value of r is 1, 6, 12, 18; is the convolution kernel weight with a dilation rate of r, is the bias with void rate r, and for Convolution kernel weights and biases.

[0100] At the same time, the feature map The feature map and primary feature map obtained after the linear layer and Relu activation function layer Perform channel splicing, and then enter the SSM-ASPP module again to extract multi-scale spatial features to obtain the intermediate feature map ; Then the feature map With feature map Channel splicing The convolution layer linearly combines the features of each channel, and then passes through the linear layer and the Relu activation function layer and combines them with the intermediate feature map Perform channel splicing and enter the SSM-ASPP module to extract multi-scale spatial features to obtain high-level feature maps , the obtained primary feature map , the intermediate feature map With high-level feature maps Perform channel splicing and then pass The convolutional layer, the linear layer and the Relu activation function layer integrate the spatial features to obtain the final global spatial feature map.

[0101] Specifically, in this embodiment, step three constructs the R-TCN module and performs time feature extraction, such as Figure 3 As shown, including:

[0102] Construct an N-layer R-TCN network module. First, use causal hole pyramid convolution to model the time series and extract features in each layer of the R-TCN network module. The formula for extracting time features by causal hole pyramid convolution is as follows:

[0103] ;

[0104] ;

[0105] in, Representative Output features of frame images, Indicates Input features of frame images, Indicates A dilated convolution kernel, express Atrous convolution kernel, Represents the convolution kernel Middle position The weight value of represents the dilation rate of the d-th dilated convolution, which determines the temporal feature extraction interval; In terms of time dimension, The step size of jump forward is used to extract features, such as , then extract the features between adjacent frame images, such as , then extract the features between the current frame image and the previous two frames of image, and so on.

[0106] In this embodiment, the TCN module constructed uses 4 parallel The causal hollow pyramid convolution kernel of the three convolution kernels has a hollow rate of 1, 2, 3, and 4 respectively. The values ​​of are 1, 2, 3, 4, , The values ​​are 1, 2, 3, and 4.

[0107] Finally, the temporal features calculated by each dilated convolution kernel are , through the splicing function Concatenate to get the final output features.

[0108] Input feature map The temporal feature map is obtained through the causal hole pyramid convolution layer , the feature map After the batch normalization layer, the Swish activation function layer and the Stochastic Depth regularization mechanism layer, the feature map is obtained. , the formulas are as follows:

[0109] ;

[0110] ;

[0111] ;

[0112] in, and are the mean and variance within the batch, respectively. and are the scaling and translation parameters, is an adjustable parameter. is a random variable with a value of 0 or 1. is a constant used to avoid division by 0 during batch normalization. is the activation function.

[0113] The nonlinear transformation capability and stability of image features are improved through the Swish activation function layer and the Stochastic Depth regularization mechanism layer. Through residual connection and feature map Add to get the feature map As the output of the first layer TCN network module.

[0114] The time feature map obtained by the first layer TCN network module As input, it is put into the second-layer TCN network module to obtain the feature map , and then the feature map Through residual connection and feature map After addition, it is put into the third-layer TCN network module as input. The whole process is repeated N-1 times. Finally, the output of the N-1th layer TCN network module is added to the output of the Nth layer TCN network module through a residual connection to obtain the final global temporal feature map of the input image.

[0115] Specifically, in this embodiment, step 4 constructs the ITSAM module and performs feature processing on the spatiotemporal feature graph, such as Figure 4 As shown, including:

[0116] The ITSAM module contains three sub-modules and a multi-head attention mechanism. The three sub-modules are INCM (embedding module), SAM (spatial attention module) and TAM (temporal attention module). After the temporal feature map and spatial feature map of step 2 and step 3 are concatenated, they will enter INCM for feature fusion.

[0117] INCM Figure 8 As shown in Figure 2, the spatiotemporal feature map first enters four parallel depth-wise separable convolutional layers, and each path extracts features at different scales: the first path uses Depthwise separable convolution extracts raw spatiotemporal features; the second path first passes through Depth-wise separable convolution, then Depthwise separable convolution to capture local spatiotemporal features; the third path uses and 5 Depth-wise separable convolution to extract local spatiotemporal features over a larger range; the fourth path is through The pooling layer obtains global spatiotemporal feature information, followed by a The depthwise separable convolution adjusts the number of channels. The output feature maps of the first three paths are combined through a splicing operation to form a multi-channel comprehensive local spatiotemporal feature map. The spliced ​​feature map is then passed through a Depthwise separable convolution further fuses information and reduces the number of redundant channels. Finally, the fused spatiotemporal feature map is locally and globally fused with the global spatiotemporal features obtained by the fourth path to generate a spatiotemporal feature map containing multi-scale information.

[0118] The spatiotemporal feature maps processed by INCM are input into TAM and SAM respectively. Figure 6 As shown in Figure 2, the fused feature map output by INCM first undergoes convolution of four parallel paths to extract features of different scales. The first path uses Convolution extracts basic features, and the second path passes through a Convolution and Convolution is used to capture local features, and the third path uses two consecutive Convolution extracts mid-scale features. The fourth path uses two consecutive 5 convolutions capture a wider range of features. Then, the last two paths are fused by feature multiplication, and the fused features are fused again with the feature map of the second path, and normalized by two Softmax functions to generate two sets of spatial attention weights. Subsequently, these two sets of weights are multiplied and fused into the final spatial attention weights, which are applied to the basic features extracted by the first path to obtain the enhanced output feature map. The entire SAM module is from input feature X to output feature X. The formula is as follows:

[0119] ;

[0120] ;

[0121] in, is the feature map obtained by convolution of the second, third, and fourth paths. Represents the fused spatial attention weight.

[0122] TAM Figure 7 As shown in , the input fused spatiotemporal feature map is first downsampled to reduce the spatial dimension, so as to focus on the extraction of global temporal features. Convolution is combined with the ReLU activation function to process the downsampled feature map to capture local and time-dependent information in the temporal features. Subsequently, the feature map is restored to its original size through a deconvolution operation to ensure that the processed temporal information is spatially aligned with the initial feature map. Next, the deconvolved feature map passes through a fully connected layer and a Sigmoid activation function to generate temporal attention weights. Finally, the generated temporal attention weights are element-wise multiplied to the original input temporal feature map to obtain the output feature map with enhanced temporal information. The entire TAM module is from input feature X to output feature X. The formula is as follows:

[0123] ;

[0124] ;

[0125] in, and is the weight matrix and bias of the fully connected layer, is the temporal attention weight, is the deconvolution function, is the downsampling function, is the bias term for deconvolution.

[0126] After being processed by TAM and SAM, the obtained feature maps are multiplied and input into the multi-head attention layer. The multiplied fusion feature map is first processed by Convolutional linear transformation generates query, key, and value matrices, which are input into multiple attention heads for processing.

[0127] For each attention head, the query and key matrices calculate the similarity between each spatial position through a dot product operation, and then the similarity result is scaled and normalized by Softmax to obtain the attention weight of the spatiotemporal position. Then, the attention weight is applied to the value matrix to achieve weighted summation to generate the output of each attention head. The outputs of these different attention heads are spliced ​​in the channel dimension to form the summary result of multi-head attention. Finally, the spliced ​​feature map is restored to the original number of channels through a linear transformation to generate the output feature map of the multi-head attention layer. The process is as follows:

[0128] ;

[0129] in, is the output feature map of the multi-head attention layer, , , is the linear transformation matrix of query, key and value, is the linear transformation matrix after multi-head output, represents the number of heads of multi-head attention, X is the input feature, is the feature dimension of the attention head, and the superscript T indicates transposition.

[0130] The multiplied fusion feature map is added to the feature map output by the multi-head attention layer through residual connection as the final output feature map. The output feature map passes through the global pooling and fully connected layer. Global pooling performs dimensionality reduction processing on the output feature map and compresses the feature information into a feature vector of a fixed length. The fully connected layer is responsible for linear mapping the feature vector and converting it into a specific predicted value, that is, the predicted value of the typhoon wind speed.

[0131] Specifically, in this embodiment, step 5 uses the Seagull Algorithm optimized by the adaptive genetic algorithm (AGA-SOA) to optimize and train the hyperparameters of the initial wind speed prediction model, such as Figure 5 As shown, including:

[0132] Step 501, construct a population consisting of different hyperparameter combinations, a population includes m hyperparameter combinations, each of which can be represented as a vector , where each element Represents the value of a hyperparameter. A hyperparameter combination has n hyperparameters, and each hyperparameter combination is an individual seagull.

[0133] Step 502: Select the inverse of the mean square error (MSE) as the adaptability function, and the formula is as follows:

[0134] ;

[0135] in, N is the sample size, is the adaptability value, For the The true value of the samples, For the The predicted value of samples, The larger the value, the better the hyperparameter combination.

[0136] Set the maximum number of iterations to , the adaptive threshold is ,Put all the seagull individuals in the population into the initial wind speed prediction model for training, and then use the adaptive function to calculate the adaptability of each seagull individual.

[0137] Step 503: Select the top population with the best adaptability value. A seagull, here Randomly select two individuals from and as a parent, then use adaptive scale Perform a linear crossover, The calculation method is as follows:

[0138] ;

[0139] in, and Parent and The adaptability value of is a constant, usually 0.5, to control the proportional adjustment amplitude. Two offspring seagull individuals are generated by linear crossover. and :

[0140] ;

[0141] ;

[0142] Step 504, the generated offspring individuals enter the mutation stage after searching. The mutation operation introduces appropriate randomness and fine-tuning, so that the offspring seagull individuals explore new solution spaces and optimize probabilities based on the parent generation. The initial mutation probability is set to , as the number of iterations increases, the current mutation probability is There will also be dynamic changes.

[0143] ;

[0144] in, is the number of iterations, It is a constant that controls the descent speed, usually 0.05-0.1.

[0145] For linear crossover, two offspring seagull individuals are generated and Each hyperparameter in A random number will be generated randomly ,when , adaptive mutation will be performed, otherwise the original value will be maintained. The process of adaptive mutation is as follows:

[0146] ;

[0147] in, is the current individual’s adaptability value, is the maximum fitness value in the population, is the maximum variation range, is a uniformly distributed random number.

[0148] Two parent seagull individuals and After two offspring individuals are generated through adaptive crossover and mutation, Randomly select two individuals from the seagull individuals and As the parent, two offspring individuals are generated through adaptive crossover and mutation. and , the whole process cycle Then all the seagull offspring individuals obtained through adaptive crossover and mutation are taken as the new generation population.

[0149] In step 505, each seagull individual in the new population will complete the migration process and the attack process on the target. The migration process simulates the behavior of the seagull approaching the target location so that the individuals in the population gradually move closer to potential excellent individuals.

[0150] First, a seagull individual is randomly selected from the new generation population. As the target location of the migration process, each individual migrates to the target location, and the new location after migration is:

[0151] ;

[0152] in, is the current location of the individual, is another individual randomly selected from the population as a disturbance term to introduce diversity, and are two control factors used to adjust the speed and direction of each individual approaching the target.

[0153] The attack process simulates the diving behavior of seagulls when hunting, and further adjusts the position of individuals so that they can conduct local searches while approaching the target position, thereby enhancing the approach to the optimal individual.

[0154] Each individual seagull that has undergone the migration process , will attack the target position, and the new position after the attack process is completed is:

[0155] ;

[0156] in, is the attack step length control factor, which is used to adjust the intensity of the attack process and ensure that the individual makes fine adjustments while approaching the target position; is a A random angle within a range, simulating the random swooping direction of a seagull when hunting.

[0157] Step 506: After each seagull individual has undergone the migration process and the attack process, a new population with better performance is generated. All individuals in the new population are put into the wind speed prediction model for training. After the training is completed, the adaptive function is used to calculate the adaptive value of each individual in the new population to determine whether the adaptive value of the best individual in the population has reached the maximum threshold. If not, the number of iterations is +1, and adaptive crossover and mutation are continued for the new population. If the adaptive value of the best individual in the population reaches the maximum threshold Or when the maximum number of iterations is reached, the optimal individual is output, and the hyperparameter combination it represents is the optimal hyperparameter combination of the model.

[0158] In step 6, the image data of the previous hour is input into the target wind speed prediction model to predict the typhoon wind speed of the next hour, including:

[0159] The optimal hyperparameter combination is put into the wind speed prediction model for training. After the training is completed, the optimal weight of the model is obtained. The target wind speed prediction model is constructed using the optimal weight to perform typhoon wind speed prediction.

[0160] The satellite cloud image data of the previous hour is put into the target wind speed prediction model to predict the cloud image feature map of the next hour. If the size of the cloud image feature map is , the number of channels is , pass it through a fully connected layer to get the typhoon wind speed forecast for the next hour , the formula is as follows:

[0161] ;

[0162] in, is the first feature map in the fully connected layer. The weight value of each channel, In the feature map Coordinates and The element value of each channel, is the bias value of the fully connected layer.

[0163] In summary, the typhoon wind speed prediction method based on Mamba-ASPP and R-TCN in the present invention integrates characteristic information of different time and space scales to cope with the complexity of typhoon wind speed changes, thereby effectively improving the efficiency and accuracy of typhoon prediction.

[0164] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A typhoon wind speed prediction method based on Mamba-ASPP and R-TCN, characterized in that: The steps include: Step 1: Obtain a satellite cloud image dataset and preprocess the image to obtain a preprocessed image dataset; Step 2: construct a Mamba-ASPP module, input the image data of adjacent frames in the preprocessed image data set into the Mamba-ASPP module, perform spatial feature extraction, and obtain a global spatial feature map of the input image, including the following sub-steps: Step 201, constructing a Mamba-ASPP module, combining a Mamba network module and an ASPP network module, in the Mamba network module, extracting spatial features of an image through a state space module and a depthwise separable convolution, and in the ASPP network module, extracting global spatial features of an image through multi-scale dilated convolution; Step 202: The pre-processed image is input into the Mamba-ASPP module. The image is sequentially processed Depthwise separable convolutional layers, Depthwise separable convolutional layers, Depthwise separable convolutional layers and Depth-wise separable convolutional layer, resulting in four feature maps , , , ; Step 203: The feature map With feature map Perform channel splicing and pass through The convolution layer linearly combines the features of each channel, and the linear layer and Relu activation function layer enhance the nonlinear representation ability of the feature map, and then enter the SSM-ASPP module to extract multi-scale spatial features to obtain the primary feature map. ; Step 204: The feature map The feature map obtained after the linear layer and the Relu activation function layer is the same as the primary feature map Perform channel splicing and enter the SSM-ASPP module again to extract multi-scale spatial features to obtain the intermediate feature map ; Step 205: The feature map With feature map To splice channels, The convolution layer linearly combines the features of each channel, passes through the linear layer and the Relu activation function layer, and is combined with the intermediate feature map Perform channel splicing and enter the SSM-ASPP module to extract multi-scale spatial features to obtain high-level feature maps ; Step 206: Get the primary feature map , Intermediate feature map , high-level feature maps Perform channel splicing and pass through Convolutional layer, linear layer, Relu activation function layer to obtain the global spatial feature map of the input image; Step 3: construct an R-TCN module, input the image data of adjacent frames in the preprocessed image data set into the R-TCN module, extract the temporal features, and obtain the global temporal feature map of the input image, including the following sub-steps: Step 301: construct an N-layer R-TCN network module. Each layer of the R-TCN network module uses causal dilation convolution to model the time series, and fuses the dilated pyramid convolution to obtain the causal dilated pyramid convolution. For the input feature map , the temporal features are extracted through the causal dilated pyramid convolution, and the temporal features calculated by each dilated convolution kernel are spliced ​​to obtain the temporal feature map ; Step 302: Time characteristic graph Put it into the first layer of TCN network module, and after the batch normalization layer, Swish activation function layer, and Stochastic Depth regularization mechanism layer, the nonlinear transformation ability and stability of the features are improved to obtain the feature map , the input feature map Through residual connection and feature map Add to get the feature map ; Step 303: The feature map As input, it is put into the second layer TCN network module to obtain the feature map , the feature map Through residual connection and feature map After addition, it is put into the third-layer TCN network module as input. After repeating N-1 times, the output of the N-1th layer TCN network module is added to the output of the Nth layer TCN network module through a residual connection to obtain the global temporal feature map of the input image. Step 4: construct an ITSAM module, splice the global spatial feature map and the global temporal feature map of the input image, put them into the ITSAM module for comprehensive feature processing, and obtain the predicted value of the typhoon wind speed; Step 5: Integrate the constructed Mamba-ASPP module, R-TCN module, and ITSAM module to form an initial wind speed prediction model, optimize and train the hyperparameters of the initial wind speed prediction model using the Seagull algorithm optimized by the adaptive genetic algorithm, and use the trained model as the target wind speed prediction model; Step 6: Input the image data of the previous hour into the target wind speed prediction model to predict the typhoon wind speed of the next hour.

2. The typhoon wind speed prediction method based on Mamba-ASPP and R-TCN according to claim 1, characterized in that: In step 1, the image is preprocessed, including the following sub-steps: Step 101: Use Gaussian filtering to remove image noise from the acquired satellite cloud image data set. The formula is as follows: ； in, is the standard deviation of the Gaussian kernel, is the value of the Gaussian kernel, and represents the two-dimensional plane coordinates of the Gaussian kernel, The function is a natural exponential function of the form ; Step 102: Use the adaptive threshold segmentation method to perform threshold segmentation on the image. The formula is as follows: ； in, It's a window. is the number of pixels in the window, is a regulation constant, is the coordinate in the image The pixel gray value; Step 103, use 3 The principle is to identify outliers in the image and replace them with the median of adjacent pixels. The principle formula is as follows: ； in, is the standard deviation of the image pixel values, is the image mean, Represents the value of the i-th pixel in the image. If the pixel value in the image satisfies 3 Principle formula, then it is defined as an outlier; Use bilinear interpolation to handle missing values. The bilinear interpolation formula is as follows: ； in, Represents the coordinates of the interpolation target point, ,express exist and The relative position between ,express exist and The relative position between Represents the values ​​of four known adjacent pixels; Step 104: Use the Min-Max normalization method to normalize the image after bilinear interpolation, and map the pixel values ​​to the range of [0, 1]. The formula is as follows: ； in, and are the minimum and maximum values ​​of the image pixels, respectively.

3. The typhoon wind speed prediction method based on Mamba-ASPP and R-TCN according to claim 1, characterized in that: The depthwise separable convolution in step 202 includes depthwise convolution and pointwise convolution; The deep convolution extracts spatial features in each channel of the image, and the formula is as follows: ； The point convolution is used to integrate the features between each channel, and the formula is as follows: ； in, Represents image coordinates With The eigenvalues ​​at the channels, is the eigenvalue of the input image, Indicated in Channels The weight of , is the convolution kernel size, Point convolution The weight of the convolution kernel, is the number of channels after depthwise convolution, is the channel number index after point convolution, The output image is at coordinates With The spatial eigenvalues ​​at the channels.

4. The typhoon wind speed prediction method based on Mamba-ASPP and R-TCN according to claim 1, characterized in that: In step 302, the temporal features are extracted by causal dilated pyramid convolution, and the formula is as follows: ； ； in, Representative Output features of frame images, Indicates Input features of frame images, Indicates A dilated convolution kernel, express Atrous convolution kernel, Represents the convolution kernel Middle position The weight value of represents the dilation rate of the d-th dilated convolution, which determines the temporal feature extraction interval; In terms of time dimension, Jump forward with a step size of to extract features; is the concatenation function, When the value is 1, 2, 3, or 4, the temporal features calculated by each dilated convolution kernel are Splice to get the final output features .

5. The typhoon wind speed prediction method based on Mamba-ASPP and R-TCN according to claim 1, characterized in that: In step 4, the ITSAM module is constructed and comprehensive feature processing is performed, including the following sub-steps: Step 401, construct an ITSAM module, including a multi-head attention mechanism, an embedding submodule, a spatial attention submodule, and a temporal attention submodule; the spatial attention submodule is used to weight and select spatial features, and the temporal attention submodule is used to strengthen and select key information in the temporal dimension; Step 402: The global spatial feature map and the global temporal feature map of the input image are spliced ​​and input into the embedding submodule for feature fusion, and then input into the spatial attention submodule and the temporal attention submodule for processing respectively. The obtained feature maps are multiplied and input into the multi-head attention layer to fuse feature information from different dimensions. Step 403: Add the multiplied feature map to the feature map output by the multi-head attention layer through a residual connection to obtain an output feature map; Step 404: The output feature map is subjected to global pooling and a fully connected layer. The global pooling performs dimensionality reduction processing on the output feature map and compresses the feature information into a feature vector of a fixed length. The fully connected layer performs linear mapping on the feature vector and converts it into a specific predicted value, which is the predicted value of the typhoon wind speed.

6. The typhoon wind speed prediction method based on Mamba-ASPP and R-TCN according to claim 5, characterized in that: In step 5, the hyperparameters of the initial wind speed prediction model are optimized and trained using the Seagull algorithm optimized by the adaptive genetic algorithm, including the following sub-steps: Step 501: Construct a population consisting of different hyperparameter combinations, the population includes m hyperparameter combinations, each of which is represented by a vector , where each element Represents the value of a hyperparameter. A hyperparameter combination has n hyperparameters, each hyperparameter combination is a seagull individual; Step 502: Select the inverse of the mean square error as the adaptability function, and the formula is as follows: ； in, N is the sample size, is the adaptability value, For the The true value of the samples, For the The predicted value of samples, The larger the value, the better the hyperparameter combination; Set the maximum number of iterations to , the adaptive threshold is , put all seagull individuals in the population into the initial wind speed prediction model for training, and calculate the adaptability of each seagull individual through the adaptive function; Step 503: Select the top population with the best adaptability value. Seagull individuals, in Randomly select two individuals from and As a parent, use adaptive scale Perform a linear crossover, The calculation method is as follows: ； in, and Parent and The adaptability value of is a constant; Generate two offspring seagull individuals through linear crossover and : ； ； Step 504, enter the mutation phase, the mutation operation introduces appropriate randomness and fine-tuning, so that the generated offspring seagull individuals explore new solution space and optimization probability based on the parent generation; set the initial mutation probability to , as the number of iterations increases, the current mutation probability is Dynamic changes also occur: ； in, is the number of iterations, is a constant that controls the descent speed; For linear crossover, two offspring seagull individuals are generated and , each hyperparameter A random number will be generated randomly ,when , adaptive mutation is performed, otherwise the original value is maintained; the hyperparameters of the offspring seagull individuals The process of adaptive mutation is as follows: ； in, is the current individual’s adaptability value, is the maximum fitness value in the population, is the maximum variation range, is a uniformly distributed random number; Two parent seagull individuals and After two offspring individuals are generated through adaptive crossover and mutation, Randomly select two individuals from the seagull individuals and As the parent, two offspring individuals are generated through adaptive crossover and mutation. and ,cycle After that, all the seagull offspring individuals obtained through adaptive crossover and mutation are regarded as the new generation population; Step 505: Each seagull individual in the new generation population completes the migration process and the attack process to the target. The migration process simulates the behavior of the seagull approaching the target position, so that the individuals in the population gradually approach potential excellent individuals; the attack process simulates the diving behavior of the seagull when hunting, adjusts the individual position, and makes the individual conduct local search while approaching the target position, thereby enhancing the approach to the optimal individual. Step 506: After each seagull individual has gone through the migration process and the attack process, a better new population is generated. After all individuals in the new population are put into the wind speed prediction model for training, the adaptive function is used to calculate the adaptive value of each individual in the new population to determine whether the adaptive value of the best individual in the new population has reached the maximum threshold. If not, the number of iterations is increased by 1, and steps 503 to 506 are repeated to perform adaptive crossover and mutation on the new population. If the adaptive value of the best individual in the population reaches the maximum threshold Or when the maximum number of iterations is reached, the optimal individual is output and the hyperparameter combination of the optimal individual is obtained.

7. The typhoon wind speed prediction method based on Mamba-ASPP and R-TCN according to claim 6, characterized in that: In step 5, during the migration process, a seagull individual is randomly selected from the new generation population. As the target location of the migration process, each individual migrates to the target location, and the new location after migration is: ； in, is the current location of the individual, is another individual randomly selected from the population as a disturbance term to introduce diversity, and are two control factors used to adjust the speed and direction of each individual approaching the target; Each individual seagull that has undergone the migration process , attack the target position, and the new position after the attack process is completed is: ； in, is the attack step length control factor, which is used to adjust the intensity of the attack process. is A random angle within a range, simulating the random swooping direction of a seagull when hunting.

8. The typhoon wind speed prediction method based on Mamba-ASPP and R-TCN according to claim 7, characterized in that: In step 6, the image data of the previous hour is input into the target wind speed prediction model to predict the typhoon wind speed of the next hour, including the following sub-steps: Step 601: Put the obtained optimal individual hyperparameter combination into the initial wind speed prediction model for training. After the training is completed, the optimal weight of the model is obtained, and the target wind speed prediction model is constructed using the optimal weight to perform typhoon wind speed prediction; Step 602: Put the satellite cloud image data of the previous hour into the target wind speed prediction model to obtain the cloud image feature map of the next hour. If the size of the cloud image feature map is , the number of channels is , pass the cloud feature map through a fully connected layer to get the typhoon wind speed forecast value for the next hour , the formula is as follows: ； in, is the first feature map in the fully connected layer. The weight value of each channel, In the feature map Coordinates and The element value of each channel, is the bias value of the fully connected layer.

Images